[ { "id": "qwen3-0.6b", "name": "Qwen3-0.6B", "repo": "Qwen/Qwen3-0.6B", "params_b": 0.75, "params_source": "https://huggingface.co/Qwen/Qwen3-0.6B", "active_params_b": null, "active_source": null, "q4_file_gb": 0.48, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Qwen_Qwen3-0.6B-GGUF", "accessed": "2026-10-15", "status": "older", "superseded_by": "qwen3.5-4b", "kv": { "layers": 28, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 40960, "config_source": "https://huggingface.co/Qwen/Qwen3-0.6B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Qwen/Qwen3-0.6B", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/qwen3:0.6b", "runs": "local", "pull": "qwen3:0.6b" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/qwen-lm/qwen-3/transformers/0.6b" } ] }, { "id": "qwen3-4b", "name": "Qwen3-4B", "repo": "Qwen/Qwen3-4B", "params_b": 4.02, "params_source": "https://huggingface.co/Qwen/Qwen3-4B", "active_params_b": null, "active_source": null, "q4_file_gb": 2.5, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Qwen_Qwen3-4B-GGUF", "accessed": "2026-10-15", "status": "older", "superseded_by": "qwen3.5-4b", "kv": { "layers": 36, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 40960, "config_source": "https://huggingface.co/Qwen/Qwen3-4B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Qwen/Qwen3-4B", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/qwen3:4b-q4_K_M", "runs": "local", "pull": "qwen3:4b-q4_K_M", "note": "The plain qwen3:4b tag now points to the Thinking-2507 update" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/qwen-lm/qwen-3/transformers/4b" } ] }, { "id": "qwen3-8b", "name": "Qwen3-8B", "repo": "Qwen/Qwen3-8B", "params_b": 8.19, "params_source": "https://huggingface.co/Qwen/Qwen3-8B", "active_params_b": null, "active_source": null, "q4_file_gb": 5.03, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Qwen_Qwen3-8B-GGUF", "accessed": "2026-10-15", "status": "older", "superseded_by": "qwen3.5-9b", "kv": { "layers": 36, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 40960, "config_source": "https://huggingface.co/Qwen/Qwen3-8B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Qwen/Qwen3-8B", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/qwen3:8b", "runs": "local", "pull": "qwen3:8b" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/qwen-lm/qwen-3/transformers/8b" } ] }, { "id": "qwen3-14b", "name": "Qwen3-14B", "repo": "Qwen/Qwen3-14B", "params_b": 14.77, "params_source": "https://huggingface.co/Qwen/Qwen3-14B", "active_params_b": null, "active_source": null, "q4_file_gb": 9.0, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Qwen_Qwen3-14B-GGUF", "accessed": "2026-10-15", "status": "older", "superseded_by": "qwen3.5-9b", "kv": { "layers": 40, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 40960, "config_source": "https://huggingface.co/Qwen/Qwen3-14B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Qwen/Qwen3-14B", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/qwen3:14b", "runs": "local", "pull": "qwen3:14b" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/qwen-lm/qwen-3/transformers/14b" } ] }, { "id": "qwen3-32b", "name": "Qwen3-32B", "repo": "Qwen/Qwen3-32B", "params_b": 32.76, "params_source": "https://huggingface.co/Qwen/Qwen3-32B", "active_params_b": null, "active_source": null, "q4_file_gb": 19.76, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Qwen_Qwen3-32B-GGUF", "accessed": "2026-10-15", "status": "older", "superseded_by": "qwen3.6-27b", "kv": { "layers": 64, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 40960, "config_source": "https://huggingface.co/Qwen/Qwen3-32B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Qwen/Qwen3-32B", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/qwen3:32b", "runs": "local", "pull": "qwen3:32b" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/qwen-lm/qwen-3/transformers/32b" } ] }, { "id": "qwen3-30b-a3b", "name": "Qwen3-30B-A3B", "repo": "Qwen/Qwen3-30B-A3B", "params_b": 30.53, "params_source": "https://huggingface.co/Qwen/Qwen3-30B-A3B", "active_params_b": 3.3, "active_source": "https://huggingface.co/Qwen/Qwen3-30B-A3B", "q4_file_gb": 18.63, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Qwen_Qwen3-30B-A3B-GGUF", "accessed": "2026-10-15", "status": "older", "superseded_by": "qwen3.6-35b-a3b", "kv": { "layers": 48, "kv_heads": 4, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 40960, "config_source": "https://huggingface.co/Qwen/Qwen3-30B-A3B/blob/main/config.json" }, "moe": { "expert_b": 28.991, "expert_active_b": 1.812, "layers": 48, "experts": 128, "experts_per_tok": 8, "expert_width": 768, "config_source": "https://huggingface.co/Qwen/Qwen3-30B-A3B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Qwen/Qwen3-30B-A3B", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/qwen3:30b-a3b-q4_K_M", "runs": "local", "pull": "qwen3:30b-a3b-q4_K_M", "note": "The plain qwen3:30b-a3b tag now points to the Thinking-2507 update" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/qwen-lm/qwen-3/transformers/30b-a3b" } ] }, { "id": "qwen3-235b-a22b", "name": "Qwen3-235B-A22B", "repo": "Qwen/Qwen3-235B-A22B", "params_b": 235.09, "params_source": "https://huggingface.co/Qwen/Qwen3-235B-A22B", "active_params_b": 22, "active_source": "https://huggingface.co/Qwen/Qwen3-235B-A22B", "q4_file_gb": 142.15, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/unsloth/Qwen3-235B-A22B-GGUF", "accessed": "2026-10-16", "kv": { "layers": 94, "kv_heads": 4, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 40960, "config_source": "https://huggingface.co/Qwen/Qwen3-235B-A22B/blob/main/config.json" }, "moe": { "expert_b": 227.096, "expert_active_b": 14.194, "layers": 94, "experts": 128, "experts_per_tok": 8, "expert_width": 1536, "config_source": "https://huggingface.co/Qwen/Qwen3-235B-A22B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Qwen/Qwen3-235B-A22B", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/qwen3:235b-a22b-q4_K_M", "runs": "local", "pull": "qwen3:235b-a22b-q4_K_M", "note": "The plain qwen3:235b-a22b tag now points to the Thinking-2507 update" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/qwen-lm/qwen-3/transformers/235b-a22b" } ] }, { "id": "qwen3-coder-30b-a3b-instruct", "name": "Qwen3-Coder-30B-A3B-Instruct", "repo": "Qwen/Qwen3-Coder-30B-A3B-Instruct", "params_b": 30.53, "params_source": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct", "active_params_b": 3.3, "active_source": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct", "q4_file_gb": 18.56, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF", "accessed": "2026-10-16", "status": "older", "superseded_by": "qwen3-coder-next", "kv": { "layers": 48, "kv_heads": 4, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 262144, "config_source": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct/blob/main/config.json" }, "moe": { "expert_b": 28.991, "expert_active_b": 1.812, "layers": 48, "experts": 128, "experts_per_tok": 8, "expert_width": 768, "config_source": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Qwen/Qwen3-Coder-30B-A3B-Instruct", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/qwen3-coder:30b", "runs": "local", "pull": "qwen3-coder:30b" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/qwen-lm/qwen3-coder/transformers/30b-a3b-instruct" } ] }, { "id": "qwen2.5-coder-7b-instruct", "name": "Qwen2.5-Coder-7B-Instruct", "repo": "Qwen/Qwen2.5-Coder-7B-Instruct", "params_b": 7.62, "params_source": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct", "active_params_b": null, "active_source": null, "q4_file_gb": 4.68, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Qwen2.5-Coder-7B-Instruct-GGUF", "accessed": "2026-10-15", "status": "older", "superseded_by": "qwen3.5-9b", "kv": { "layers": 28, "kv_heads": 4, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 32768, "config_source": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Qwen/Qwen2.5-Coder-7B-Instruct", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/qwen2.5-coder:7b-instruct", "runs": "local", "pull": "qwen2.5-coder:7b-instruct" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/qwen-lm/qwen2.5-coder/transformers/7b-instruct" } ] }, { "id": "qwen2.5-coder-14b-instruct", "name": "Qwen2.5-Coder-14B-Instruct", "repo": "Qwen/Qwen2.5-Coder-14B-Instruct", "params_b": 14.77, "params_source": "https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct", "active_params_b": null, "active_source": null, "q4_file_gb": 8.99, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Qwen2.5-Coder-14B-Instruct-GGUF", "accessed": "2026-10-15", "status": "older", "superseded_by": "qwen3.5-9b", "kv": { "layers": 48, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 32768, "config_source": "https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Qwen/Qwen2.5-Coder-14B-Instruct", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/qwen2.5-coder:14b-instruct", "runs": "local", "pull": "qwen2.5-coder:14b-instruct" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/qwen-lm/qwen2.5-coder/transformers/14b-instruct" } ] }, { "id": "qwen2.5-coder-32b-instruct", "name": "Qwen2.5-Coder-32B-Instruct", "repo": "Qwen/Qwen2.5-Coder-32B-Instruct", "params_b": 32.76, "params_source": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct", "active_params_b": null, "active_source": null, "q4_file_gb": 19.85, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Qwen2.5-Coder-32B-Instruct-GGUF", "accessed": "2026-10-15", "status": "older", "superseded_by": "qwen3-coder-next", "kv": { "layers": 64, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 32768, "config_source": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Qwen/Qwen2.5-Coder-32B-Instruct", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/qwen2.5-coder:32b-instruct", "runs": "local", "pull": "qwen2.5-coder:32b-instruct" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/qwen-lm/qwen2.5-coder/transformers/32b-instruct" } ] }, { "id": "qwen2.5-72b-instruct", "name": "Qwen2.5-72B-Instruct", "repo": "Qwen/Qwen2.5-72B-Instruct", "params_b": 72.71, "params_source": "https://huggingface.co/Qwen/Qwen2.5-72B-Instruct", "active_params_b": null, "active_source": null, "q4_file_gb": 47.42, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF", "accessed": "2026-10-15", "status": "older", "superseded_by": "qwen3.6-27b", "kv": { "layers": 80, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 32768, "config_source": "https://huggingface.co/Qwen/Qwen2.5-72B-Instruct/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Qwen/Qwen2.5-72B-Instruct", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/qwen2.5:72b-instruct", "runs": "local", "pull": "qwen2.5:72b-instruct" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/qwen-lm/qwen2.5/transformers/72b-instruct" } ] }, { "id": "llama-3.1-8b-instruct", "name": "Llama-3.1-8B-Instruct", "repo": "meta-llama/Llama-3.1-8B-Instruct", "params_b": 8.03, "params_source": "https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct", "active_params_b": null, "active_source": null, "q4_file_gb": 4.92, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF", "accessed": "2026-10-15", "status": "older", "superseded_by": null, "kv": { "layers": 32, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 131072, "config_source": "https://huggingface.co/unsloth/Llama-3.1-8B-Instruct/blob/main/config.json", "config_note": "official repo is gated; config read from ungated mirror" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/AI-ModelScope/Llama-3.1-8B-Instruct", "uploader": "other", "note": "ModelScope community mirror account, not run by Meta" }, { "source": "ollama", "url": "https://ollama.com/library/llama3.1:8b", "runs": "local", "pull": "llama3.1:8b" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/metaresearch/llama-3.1/transformers/8b-instruct" }, { "source": "official", "url": "https://www.llama.com/llama-downloads/", "note": "Meta's own download request page" } ] }, { "id": "llama-3.2-3b-instruct", "name": "Llama-3.2-3B-Instruct", "repo": "meta-llama/Llama-3.2-3B-Instruct", "params_b": 3.21, "params_source": "https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct", "active_params_b": null, "active_source": null, "q4_file_gb": 2.02, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Llama-3.2-3B-Instruct-GGUF", "accessed": "2026-10-15", "status": "older", "superseded_by": null, "kv": { "layers": 28, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 131072, "config_source": "https://huggingface.co/unsloth/Llama-3.2-3B-Instruct/blob/main/config.json", "config_note": "official repo is gated; config read from ungated mirror" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/LLM-Research/Llama-3.2-3B-Instruct", "uploader": "other", "note": "ModelScope community mirror account, not run by Meta" }, { "source": "ollama", "url": "https://ollama.com/library/llama3.2:3b", "runs": "local", "pull": "llama3.2:3b" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/metaresearch/llama-3.2/transformers/3b-instruct" }, { "source": "official", "url": "https://www.llama.com/llama-downloads/", "note": "Meta's own download request page" } ] }, { "id": "llama-3.3-70b-instruct", "name": "Llama-3.3-70B-Instruct", "repo": "meta-llama/Llama-3.3-70B-Instruct", "params_b": 70.55, "params_source": "https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct", "active_params_b": null, "active_source": null, "q4_file_gb": 42.52, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Llama-3.3-70B-Instruct-GGUF", "accessed": "2026-10-15", "status": "older", "superseded_by": "muse-glimmer-30b", "kv": { "layers": 80, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 131072, "config_source": "https://huggingface.co/unsloth/Llama-3.3-70B-Instruct/blob/main/config.json", "config_note": "official repo is gated; config read from ungated mirror" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/LLM-Research/Llama-3.3-70B-Instruct", "uploader": "other", "note": "ModelScope community mirror account, not run by Meta" }, { "source": "ollama", "url": "https://ollama.com/library/llama3.3:70b", "runs": "local", "pull": "llama3.3:70b" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/metaresearch/llama-3.3/transformers/70b-instruct" }, { "source": "official", "url": "https://www.llama.com/llama-downloads/", "note": "Meta's own download request page" } ] }, { "id": "gemma-3-4b-it", "name": "gemma-3-4b-it", "repo": "google/gemma-3-4b-it", "params_b": 4.3, "params_source": "https://huggingface.co/google/gemma-3-4b-it", "active_params_b": null, "active_source": null, "q4_file_gb": 2.49, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/google_gemma-3-4b-it-GGUF", "accessed": "2026-10-15", "status": "older", "superseded_by": "gemma-4-e4b-it", "kv": { "layers": 5, "kv_heads": 4, "head_dim": 256, "v_dim": 256, "swa_layers": 29, "swa_kv_heads": 4, "swa_head_dim": 256, "window": 1024, "arch": "swa", "max_ctx": 131072, "config_source": "https://huggingface.co/unsloth/gemma-3-4b-it/blob/main/config.json", "config_note": "official repo is gated; config read from ungated mirror" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/google/gemma-3-4b-it", "uploader": "other", "note": "Account named Google; we found no Google page linking to it" }, { "source": "ollama", "url": "https://ollama.com/library/gemma3:4b", "runs": "local", "pull": "gemma3:4b" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/google/gemma-3/transformers/gemma-3-4b-it" } ] }, { "id": "gemma-3-12b-it", "name": "gemma-3-12b-it", "repo": "google/gemma-3-12b-it", "params_b": 12.19, "params_source": "https://huggingface.co/google/gemma-3-12b-it", "active_params_b": null, "active_source": null, "q4_file_gb": 7.3, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/google_gemma-3-12b-it-GGUF", "accessed": "2026-10-15", "status": "older", "superseded_by": "gemma-4-12b-it", "kv": { "layers": 8, "kv_heads": 8, "head_dim": 256, "v_dim": 256, "swa_layers": 40, "swa_kv_heads": 8, "swa_head_dim": 256, "window": 1024, "arch": "swa", "max_ctx": 131072, "config_source": "https://huggingface.co/unsloth/gemma-3-12b-it/blob/main/config.json", "config_note": "official repo is gated; config read from ungated mirror" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/google/gemma-3-12b-it", "uploader": "other", "note": "Account named Google; we found no Google page linking to it" }, { "source": "ollama", "url": "https://ollama.com/library/gemma3:12b", "runs": "local", "pull": "gemma3:12b" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/google/gemma-3/transformers/gemma-3-12b-it" } ] }, { "id": "gemma-3-27b-it", "name": "gemma-3-27b-it", "repo": "google/gemma-3-27b-it", "params_b": 27.43, "params_source": "https://huggingface.co/google/gemma-3-27b-it", "active_params_b": null, "active_source": null, "q4_file_gb": 16.55, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/google_gemma-3-27b-it-GGUF", "accessed": "2026-10-15", "status": "older", "superseded_by": "gemma-4-31b-it", "kv": { "layers": 10, "kv_heads": 16, "head_dim": 128, "v_dim": 128, "swa_layers": 52, "swa_kv_heads": 16, "swa_head_dim": 128, "window": 1024, "arch": "swa", "max_ctx": 131072, "config_source": "https://huggingface.co/unsloth/gemma-3-27b-it/blob/main/config.json", "config_note": "official repo is gated; config read from ungated mirror" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/google/gemma-3-27b-it", "uploader": "other", "note": "Account named Google; we found no Google page linking to it" }, { "source": "ollama", "url": "https://ollama.com/library/gemma3:27b", "runs": "local", "pull": "gemma3:27b" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/google/gemma-3/transformers/gemma-3-27b-it" } ] }, { "id": "mistral-7b-instruct-v0.3", "name": "Mistral-7B-Instruct-v0.3", "repo": "mistralai/Mistral-7B-Instruct-v0.3", "params_b": 7.25, "params_source": "https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3", "active_params_b": null, "active_source": null, "q4_file_gb": 4.37, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Mistral-7B-Instruct-v0.3-GGUF", "accessed": "2026-10-15", "status": "older", "superseded_by": null, "kv": { "layers": 32, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 32768, "config_source": "https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/mistralai/Mistral-7B-Instruct-v0.3", "uploader": "other", "note": "Account named Mistral AI; we found no Mistral page linking to it" }, { "source": "ollama", "url": "https://ollama.com/library/mistral:7b-instruct-v0.3-q4_K_M", "runs": "local", "pull": "mistral:7b-instruct-v0.3-q4_K_M" }, { "source": "official", "url": "https://models.mistralcdn.com/mistral-7b-v0-3/mistral-7B-Instruct-v0.3.tar", "note": "Mistral direct download (.tar), md5 listed in the mistral-inference README" } ] }, { "id": "mistral-nemo-instruct-2407", "name": "Mistral-Nemo-Instruct-2407", "repo": "mistralai/Mistral-Nemo-Instruct-2407", "params_b": 12.25, "params_source": "https://huggingface.co/mistralai/Mistral-Nemo-Instruct-2407", "active_params_b": null, "active_source": null, "q4_file_gb": 7.48, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Mistral-Nemo-Instruct-2407-GGUF", "accessed": "2026-10-15", "status": "older", "superseded_by": null, "kv": { "layers": 40, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 131072, "config_source": "https://huggingface.co/mistralai/Mistral-Nemo-Instruct-2407/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/mistralai/Mistral-Nemo-Instruct-2407", "uploader": "other", "note": "Account named Mistral AI; we found no Mistral page linking to it" }, { "source": "ollama", "url": "https://ollama.com/library/mistral-nemo:12b", "runs": "local", "pull": "mistral-nemo:12b" }, { "source": "official", "url": "https://models.mistralcdn.com/mistral-nemo-2407/mistral-nemo-instruct-2407.tar", "note": "Mistral direct download (.tar), md5 listed in the mistral-inference README" } ] }, { "id": "mistral-small-3.2-24b-instruct-2506", "name": "Mistral-Small-3.2-24B-Instruct-2506", "repo": "mistralai/Mistral-Small-3.2-24B-Instruct-2506", "params_b": 24.01, "params_source": "https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506", "active_params_b": null, "active_source": null, "q4_file_gb": 14.33, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/mistralai_Mistral-Small-3.2-24B-Instruct-2506-GGUF", "accessed": "2026-10-15", "kv": { "layers": 40, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 131072, "config_source": "https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/mistralai/Mistral-Small-3.2-24B-Instruct-2506", "uploader": "other", "note": "Account named Mistral AI; we found no Mistral page linking to it" }, { "source": "ollama", "url": "https://ollama.com/library/mistral-small3.2:24b", "runs": "local", "pull": "mistral-small3.2:24b" } ] }, { "id": "devstral-small-2505", "name": "Devstral-Small-2505", "repo": "mistralai/Devstral-Small-2505", "params_b": 23.57, "params_source": "https://huggingface.co/mistralai/Devstral-Small-2505", "active_params_b": null, "active_source": null, "q4_file_gb": 14.33, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/mistralai_Devstral-Small-2505-GGUF", "accessed": "2026-10-15", "kv": { "layers": 40, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 131072, "config_source": "https://huggingface.co/mistralai/Devstral-Small-2505/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/mistralai/Devstral-Small-2505", "uploader": "other", "note": "Account named Mistral AI; we found no Mistral page linking to it" }, { "source": "ollama", "url": "https://ollama.com/library/devstral:24b-small-2505-q4_K_M", "runs": "local", "pull": "devstral:24b-small-2505-q4_K_M" } ] }, { "id": "phi-4", "name": "phi-4", "repo": "microsoft/phi-4", "params_b": 14.66, "params_source": "https://huggingface.co/microsoft/phi-4", "active_params_b": null, "active_source": null, "q4_file_gb": 9.05, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/phi-4-GGUF", "accessed": "2026-10-16", "kv": { "layers": 40, "kv_heads": 10, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 16384, "config_source": "https://huggingface.co/microsoft/phi-4/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/microsoft/phi-4", "uploader": "other", "note": "Account labeled 'Microsoft mirror'; we found no Microsoft page linking to it" }, { "source": "ollama", "url": "https://ollama.com/library/phi4:14b", "runs": "local", "pull": "phi4:14b" } ] }, { "id": "deepseek-r1-distill-qwen-14b", "name": "DeepSeek-R1-Distill-Qwen-14B", "repo": "deepseek-ai/DeepSeek-R1-Distill-Qwen-14B", "params_b": 14.77, "params_source": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B", "active_params_b": null, "active_source": null, "q4_file_gb": 8.99, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/DeepSeek-R1-Distill-Qwen-14B-GGUF", "accessed": "2026-10-15", "status": "older", "superseded_by": null, "kv": { "layers": 48, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 131072, "config_source": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/deepseek-r1:14b-qwen-distill-q4_K_M", "runs": "local", "pull": "deepseek-r1:14b-qwen-distill-q4_K_M" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/deepseek-ai/deepseek-r1/transformers/deepseek-r1-distill-qwen-14b" } ] }, { "id": "deepseek-r1-distill-qwen-32b", "name": "DeepSeek-R1-Distill-Qwen-32B", "repo": "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B", "params_b": 32.76, "params_source": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B", "active_params_b": null, "active_source": null, "q4_file_gb": 19.85, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/DeepSeek-R1-Distill-Qwen-32B-GGUF", "accessed": "2026-10-15", "status": "older", "superseded_by": null, "kv": { "layers": 64, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 131072, "config_source": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/deepseek-r1:32b-qwen-distill-q4_K_M", "runs": "local", "pull": "deepseek-r1:32b-qwen-distill-q4_K_M" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/deepseek-ai/deepseek-r1/transformers/deepseek-r1-distill-qwen-32b" } ] }, { "id": "deepseek-r1-distill-llama-70b", "name": "DeepSeek-R1-Distill-Llama-70B", "repo": "deepseek-ai/DeepSeek-R1-Distill-Llama-70B", "params_b": 70.55, "params_source": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B", "active_params_b": null, "active_source": null, "q4_file_gb": 42.52, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/DeepSeek-R1-Distill-Llama-70B-GGUF", "accessed": "2026-10-15", "status": "older", "superseded_by": null, "kv": { "layers": 80, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 131072, "config_source": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/deepseek-ai/DeepSeek-R1-Distill-Llama-70B", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/deepseek-r1:70b-llama-distill-q4_K_M", "runs": "local", "pull": "deepseek-r1:70b-llama-distill-q4_K_M" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/deepseek-ai/deepseek-r1/transformers/deepseek-r1-distill-llama-70b" } ] }, { "id": "gpt-oss-20b", "name": "gpt-oss-20b", "repo": "openai/gpt-oss-20b", "params_b": 20.91, "params_source": "https://huggingface.co/openai/gpt-oss-20b", "active_params_b": 3.6, "active_source": "https://huggingface.co/openai/gpt-oss-20b", "q4_file_gb": 11.67, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/openai_gpt-oss-20b-GGUF", "accessed": "2026-10-15", "kv": { "layers": 12, "kv_heads": 8, "head_dim": 64, "v_dim": 64, "swa_layers": 12, "swa_kv_heads": 8, "swa_head_dim": 64, "window": 128, "arch": "swa", "max_ctx": 131072, "config_source": "https://huggingface.co/openai/gpt-oss-20b/blob/main/config.json" }, "moe": { "expert_b": 19.11, "expert_active_b": 2.389, "layers": 24, "experts": 32, "experts_per_tok": 4, "expert_width": 2880, "config_source": "https://huggingface.co/openai/gpt-oss-20b/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/openai-mirror/gpt-oss-20b", "uploader": "other", "note": "Mirror account, not run by OpenAI" }, { "source": "ollama", "url": "https://ollama.com/library/gpt-oss:20b", "runs": "local", "pull": "gpt-oss:20b" } ] }, { "id": "gpt-oss-120b", "name": "gpt-oss-120b", "repo": "openai/gpt-oss-120b", "params_b": 116.83, "params_source": "https://huggingface.co/openai/gpt-oss-120b", "active_params_b": 5.1, "active_source": "https://huggingface.co/openai/gpt-oss-120b", "q4_file_gb": 62.84, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/openai_gpt-oss-120b-GGUF", "accessed": "2026-10-15", "kv": { "layers": 18, "kv_heads": 8, "head_dim": 64, "v_dim": 64, "swa_layers": 18, "swa_kv_heads": 8, "swa_head_dim": 64, "window": 128, "arch": "swa", "max_ctx": 131072, "config_source": "https://huggingface.co/openai/gpt-oss-120b/blob/main/config.json" }, "moe": { "expert_b": 114.662, "expert_active_b": 3.583, "layers": 36, "experts": 128, "experts_per_tok": 4, "expert_width": 2880, "config_source": "https://huggingface.co/openai/gpt-oss-120b/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/openai-mirror/gpt-oss-120b", "uploader": "other", "note": "Mirror account, not run by OpenAI" }, { "source": "ollama", "url": "https://ollama.com/library/gpt-oss:120b", "runs": "local", "pull": "gpt-oss:120b" } ] }, { "id": "glm-4.5-air", "name": "GLM-4.5-Air", "repo": "zai-org/GLM-4.5-Air", "params_b": 110.47, "params_source": "https://huggingface.co/zai-org/GLM-4.5-Air", "active_params_b": 12, "active_source": "https://huggingface.co/zai-org/GLM-4.5-Air", "q4_file_gb": 73.5, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/zai-org_GLM-4.5-Air-GGUF", "accessed": "2026-10-15", "status": "older", "superseded_by": "glm-4.7-flash", "kv": { "layers": 46, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 131072, "config_source": "https://huggingface.co/zai-org/GLM-4.5-Air/blob/main/config.json" }, "moe_skip": "derived active parameters (17.04B, 15.80B without embeddings) are more than 25% from the published 12B", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/ZhipuAI/GLM-4.5-Air", "uploader": "official" } ] }, { "id": "nvidia-nemotron-nano-9b-v2", "name": "NVIDIA-Nemotron-Nano-9B-v2", "repo": "nvidia/NVIDIA-Nemotron-Nano-9B-v2", "params_b": 8.89, "params_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2", "active_params_b": null, "active_source": null, "q4_file_gb": 6.53, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/nvidia_NVIDIA-Nemotron-Nano-9B-v2-GGUF", "accessed": "2026-10-15", "status": "older", "superseded_by": "nemotron-3-nano-4b", "kv": { "layers": 4, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "hybrid", "max_ctx": 131072, "config_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/nv-community/NVIDIA-Nemotron-Nano-9B-v2", "uploader": "other", "note": "NVIDIA community account; we found no NVIDIA page linking to it" } ] }, { "id": "gemma-4-e2b-it", "name": "Gemma 4 E2B", "repo": "google/gemma-4-E2B-it", "params_b": 5.12, "params_source": "https://huggingface.co/google/gemma-4-E2B-it", "active_params_b": null, "active_source": null, "q4_file_gb": 3.46, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/google_gemma-4-E2B-it-GGUF", "accessed": "2026-10-16", "license": "Apache 2.0", "context": 131072, "released": "2026-03-02", "note": "2.3B effective parameters (5.1B with embeddings) per the model card.", "kv": { "layers": 3, "kv_heads": 1, "head_dim": 512, "v_dim": 512, "swa_layers": 12, "swa_kv_heads": 1, "swa_head_dim": 256, "window": 512, "shared_kv_layers": 20, "arch": "swa", "max_ctx": 131072, "config_source": "https://huggingface.co/google/gemma-4-E2B-it/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/google/gemma-4-E2B-it", "uploader": "other", "note": "Account named Google; we found no Google page linking to it" }, { "source": "ollama", "url": "https://ollama.com/library/gemma4:e2b", "runs": "local", "pull": "gemma4:e2b" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/google/gemma-4/transformers/gemma-4-e2b-it" } ] }, { "id": "gemma-4-e4b-it", "name": "Gemma 4 E4B", "repo": "google/gemma-4-E4B-it", "params_b": 8.0, "params_source": "https://huggingface.co/google/gemma-4-E4B-it", "active_params_b": null, "active_source": null, "q4_file_gb": 4.98, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/unsloth/gemma-4-E4B-it-GGUF", "accessed": "2026-10-16", "license": "Apache 2.0", "context": 131072, "released": "2026-03-02", "note": "4.5B effective parameters (8B with embeddings) per the model card.", "kv": { "layers": 4, "kv_heads": 2, "head_dim": 512, "v_dim": 512, "swa_layers": 20, "swa_kv_heads": 2, "swa_head_dim": 256, "window": 512, "shared_kv_layers": 18, "arch": "swa", "max_ctx": 131072, "config_source": "https://huggingface.co/google/gemma-4-E4B-it/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/google/gemma-4-E4B-it", "uploader": "other", "note": "Account named Google; we found no Google page linking to it" }, { "source": "ollama", "url": "https://ollama.com/library/gemma4:e4b", "runs": "local", "pull": "gemma4:e4b" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/google/gemma-4/transformers/gemma-4-e4b-it" } ] }, { "id": "gemma-4-12b-it", "name": "Gemma 4 12B", "repo": "google/gemma-4-12B-it", "params_b": 11.96, "params_source": "https://huggingface.co/google/gemma-4-12B-it", "active_params_b": null, "active_source": null, "q4_file_gb": 7.12, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/unsloth/gemma-4-12b-it-GGUF", "accessed": "2026-10-16", "license": "Apache 2.0", "context": 262144, "released": "2026-05-23", "kv": { "layers": 8, "kv_heads": 1, "head_dim": 512, "v_dim": 512, "swa_layers": 40, "swa_kv_heads": 8, "swa_head_dim": 256, "window": 1024, "arch": "swa", "max_ctx": 262144, "config_source": "https://huggingface.co/google/gemma-4-12B-it/blob/main/config.json", "note": "config sets attention_k_eq_v; we still count separate K and V (conservative)" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/google/gemma-4-12B-it", "uploader": "other", "note": "Account named Google; we found no Google page linking to it" }, { "source": "ollama", "url": "https://ollama.com/library/gemma4:12b", "runs": "local", "pull": "gemma4:12b" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/google/gemma-4/transformers/gemma-4-12b-it" } ] }, { "id": "gemma-4-26b-a4b-it", "name": "Gemma 4 26B A4B", "repo": "google/gemma-4-26B-A4B-it", "params_b": 25.81, "params_source": "https://huggingface.co/google/gemma-4-26B-A4B-it", "active_params_b": 3.8, "active_source": "https://huggingface.co/google/gemma-4-26B-A4B-it", "q4_file_gb": 17.04, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/google_gemma-4-26B-A4B-it-GGUF", "accessed": "2026-10-16", "license": "Apache 2.0", "context": 262144, "released": "2026-03-11", "kv": { "layers": 5, "kv_heads": 2, "head_dim": 512, "v_dim": 512, "swa_layers": 25, "swa_kv_heads": 8, "swa_head_dim": 256, "window": 1024, "arch": "swa", "max_ctx": 262144, "config_source": "https://huggingface.co/google/gemma-4-26B-A4B-it/blob/main/config.json", "note": "config sets attention_k_eq_v; we still count separate K and V (conservative)" }, "moe": { "expert_b": 22.838, "expert_active_b": 1.427, "layers": 30, "experts": 128, "experts_per_tok": 8, "expert_width": 704, "config_source": "https://huggingface.co/google/gemma-4-26B-A4B-it/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/google/gemma-4-26B-A4B-it", "uploader": "other", "note": "Account named Google; we found no Google page linking to it" }, { "source": "ollama", "url": "https://ollama.com/library/gemma4:26b-a4b", "runs": "local", "pull": "gemma4:26b-a4b" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/google/gemma-4/transformers/gemma-4-26b-a4b-it" } ] }, { "id": "gemma-4-31b-it", "name": "Gemma 4 31B", "repo": "google/gemma-4-31B-it", "params_b": 31.27, "params_source": "https://huggingface.co/google/gemma-4-31B-it", "active_params_b": null, "active_source": null, "q4_file_gb": 18.32, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/unsloth/gemma-4-31B-it-GGUF", "accessed": "2026-10-16", "license": "Apache 2.0", "context": 262144, "released": "2026-03-11", "kv": { "layers": 10, "kv_heads": 4, "head_dim": 512, "v_dim": 512, "swa_layers": 50, "swa_kv_heads": 16, "swa_head_dim": 256, "window": 1024, "arch": "swa", "max_ctx": 262144, "config_source": "https://huggingface.co/google/gemma-4-31B-it/blob/main/config.json", "note": "config sets attention_k_eq_v; we still count separate K and V (conservative)" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/google/gemma-4-31B-it", "uploader": "other", "note": "Account named Google; we found no Google page linking to it" }, { "source": "ollama", "url": "https://ollama.com/library/gemma4:31b", "runs": "local", "pull": "gemma4:31b" }, { "source": "kaggle", "url": "https://www.kaggle.com/models/google/gemma-4/transformers/gemma-4-31b-it" } ] }, { "id": "qwen3.5-4b", "name": "Qwen3.5-4B", "repo": "Qwen/Qwen3.5-4B", "params_b": 4.66, "params_source": "https://huggingface.co/Qwen/Qwen3.5-4B", "active_params_b": null, "active_source": null, "q4_file_gb": 3.01, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Qwen_Qwen3.5-4B-GGUF", "accessed": "2026-10-16", "license": "Apache 2.0", "context": 262144, "released": "2026-02-27", "kv": { "layers": 8, "kv_heads": 4, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/Qwen/Qwen3.5-4B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Qwen/Qwen3.5-4B", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/qwen3.5:4b", "runs": "local", "pull": "qwen3.5:4b" } ] }, { "id": "qwen3.5-9b", "name": "Qwen3.5-9B", "repo": "Qwen/Qwen3.5-9B", "params_b": 9.65, "params_source": "https://huggingface.co/Qwen/Qwen3.5-9B", "active_params_b": null, "active_source": null, "q4_file_gb": 6.17, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Qwen_Qwen3.5-9B-GGUF", "accessed": "2026-10-16", "license": "Apache 2.0", "context": 262144, "released": "2026-02-27", "kv": { "layers": 8, "kv_heads": 4, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/Qwen/Qwen3.5-9B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Qwen/Qwen3.5-9B", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/qwen3.5:9b", "runs": "local", "pull": "qwen3.5:9b" } ] }, { "id": "qwen3.5-27b", "name": "Qwen3.5-27B", "repo": "Qwen/Qwen3.5-27B", "params_b": 27.78, "params_source": "https://huggingface.co/Qwen/Qwen3.5-27B", "active_params_b": null, "active_source": null, "q4_file_gb": 17.98, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Qwen_Qwen3.5-27B-GGUF", "accessed": "2026-10-16", "license": "Apache 2.0", "context": 262144, "released": "2026-02-24", "kv": { "layers": 16, "kv_heads": 4, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/Qwen/Qwen3.5-27B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Qwen/Qwen3.5-27B", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/qwen3.5:27b", "runs": "local", "pull": "qwen3.5:27b" } ] }, { "id": "qwen3.5-35b-a3b", "name": "Qwen3.5-35B-A3B", "repo": "Qwen/Qwen3.5-35B-A3B", "params_b": 35.95, "params_source": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B", "active_params_b": 3, "active_source": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B", "q4_file_gb": 22.29, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Qwen_Qwen3.5-35B-A3B-GGUF", "accessed": "2026-10-16", "license": "Apache 2.0", "context": 262144, "released": "2026-02-24", "kv": { "layers": 10, "kv_heads": 2, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B/blob/main/config.json" }, "moe": { "expert_b": 32.212, "expert_active_b": 1.007, "layers": 40, "experts": 256, "experts_per_tok": 8, "expert_width": 512, "config_source": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Qwen/Qwen3.5-35B-A3B", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/qwen3.5:35b-a3b", "runs": "local", "pull": "qwen3.5:35b-a3b" } ] }, { "id": "qwen3.6-27b", "name": "Qwen3.6-27B", "repo": "Qwen/Qwen3.6-27B", "params_b": 27.78, "params_source": "https://huggingface.co/Qwen/Qwen3.6-27B", "active_params_b": null, "active_source": null, "q4_file_gb": 17.98, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Qwen_Qwen3.6-27B-GGUF", "accessed": "2026-10-16", "license": "Apache 2.0", "context": 262144, "released": "2026-04-21", "kv": { "layers": 16, "kv_heads": 4, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/Qwen/Qwen3.6-27B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Qwen/Qwen3.6-27B", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/qwen3.6:27b", "runs": "local", "pull": "qwen3.6:27b" } ] }, { "id": "qwen3.6-35b-a3b", "name": "Qwen3.6-35B-A3B", "repo": "Qwen/Qwen3.6-35B-A3B", "params_b": 35.95, "params_source": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B", "active_params_b": 3, "active_source": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B", "q4_file_gb": 22.29, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Qwen_Qwen3.6-35B-A3B-GGUF", "accessed": "2026-10-16", "license": "Apache 2.0", "context": 262144, "released": "2026-04-15", "kv": { "layers": 10, "kv_heads": 2, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/blob/main/config.json" }, "moe": { "expert_b": 32.212, "expert_active_b": 1.007, "layers": 40, "experts": 256, "experts_per_tok": 8, "expert_width": 512, "config_source": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Qwen/Qwen3.6-35B-A3B", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/qwen3.6:35b-a3b", "runs": "local", "pull": "qwen3.6:35b-a3b" } ] }, { "id": "qwen3.8-27b", "name": "Qwen3.8-27B", "repo": "Qwen/Qwen3.8-27B", "params_b": 27.78, "params_source": "https://huggingface.co/Qwen/Qwen3.8-27B", "active_params_b": null, "active_source": null, "q4_file_gb": 17.44, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Qwen3.8-27B-GGUF", "accessed": "2026-10-16", "license": "Apache 2.0", "context": 262144, "released": "2026-08-05", "kv": { "layers": 16, "kv_heads": 4, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/Qwen/Qwen3.8-27B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Qwen/Qwen3.8-27B", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/qwen3.8:27b", "runs": "local", "pull": "qwen3.8:27b" } ] }, { "id": "qwen3.8-flash-next", "name": "Qwen3.8-Flash-Next", "repo": "Qwen/Qwen3.8-Flash-Next", "params_b": 180.0, "params_source": "https://huggingface.co/Qwen/Qwen3.8-Flash-Next", "active_params_b": 6, "active_source": "https://huggingface.co/Qwen/Qwen3.8-Flash-Next", "q4_file_gb": 119.6, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Qwen3.8-Flash-Next-GGUF", "accessed": "2026-10-16", "license": "Qwen Community License 1.0", "context": 262144, "released": "2026-08-24", "note": "Model card: 125B parameters with 6B activated, plus 51B n-gram embedding and 4B MTP.", "kv": { "layers": 12, "kv_heads": 2, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/Qwen/Qwen3.8-Flash-Next/blob/main/config.json" }, "moe_skip": "derived active parameters (61.56B, 60.29B without embeddings) are more than 25% from the published 6B", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Qwen/Qwen3.8-Flash-Next", "uploader": "official" } ] }, { "id": "qwen3-coder-next", "name": "Qwen3-Coder-Next", "repo": "Qwen/Qwen3-Coder-Next", "params_b": 79.67, "params_source": "https://huggingface.co/Qwen/Qwen3-Coder-Next", "active_params_b": 3, "active_source": "https://huggingface.co/Qwen/Qwen3-Coder-Next", "q4_file_gb": 48.53, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/unsloth/Qwen3-Coder-Next-GGUF", "accessed": "2026-10-16", "license": "Apache 2.0", "context": 262144, "released": "2026-01-30", "kv": { "layers": 12, "kv_heads": 2, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/Qwen/Qwen3-Coder-Next/blob/main/config.json" }, "moe": { "expert_b": 77.309, "expert_active_b": 1.51, "layers": 48, "experts": 512, "experts_per_tok": 10, "expert_width": 512, "config_source": "https://huggingface.co/Qwen/Qwen3-Coder-Next/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Qwen/Qwen3-Coder-Next", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/qwen3-coder-next:q4_K_M", "runs": "local", "pull": "qwen3-coder-next:q4_K_M" } ] }, { "id": "muse-glimmer-30b", "name": "Meta Muse Glimmer 30B", "repo": "meta-models/Muse-Glimmer-30B", "params_b": 29.78, "params_source": "https://huggingface.co/meta-models/Muse-Glimmer-30B", "active_params_b": null, "active_source": null, "q4_file_gb": 18.39, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/meta-models/Muse-Glimmer-30B-GGUF", "accessed": "2026-10-16", "license": "Apache 2.0", "context": 131072, "released": "2026-08-09", "note": "Official GGUF from Meta. Parameter count includes the vision encoder.", "kv": { "layers": 13, "kv_heads": 2, "head_dim": 128, "v_dim": 128, "swa_layers": 39, "swa_kv_heads": 2, "swa_head_dim": 128, "window": 2048, "arch": "swa", "max_ctx": 131072, "config_source": "https://huggingface.co/meta-models/Muse-Glimmer-30B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/meta-models/Muse-Glimmer-30B", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" }, { "source": "ollama", "url": "https://ollama.com/library/muse-glimmer:30b", "runs": "local", "pull": "muse-glimmer:30b" } ] }, { "id": "muse-spark-open-816b-a42b", "name": "Meta Muse Spark Open 816B-A42B", "repo": "meta-models/Muse-Spark-Open-816B-A42B", "params_b": 822.04, "params_source": "https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B", "active_params_b": 42, "active_source": "https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B", "q4_file_gb": 505.5, "quant": "UD-Q4_K_XL", "q4_source": "https://huggingface.co/unsloth/Muse-Spark-Open-816B-A42B-GGUF", "accessed": "2026-10-16", "license": "Muse Model License (custom)", "context": null, "released": "2026-10-15", "released_source": "https://research.meta.ai/blog/muse-spark-open", "note": "Gated repo. Active params from the model name; release date from the Hugging Face model card and Meta's announcement. No plain Q4_K_M GGUF found, so the size is unsloth UD-Q4_K_XL.", "kv": { "layers": 16, "kv_heads": 16, "head_dim": 64, "v_dim": 64, "swa_layers": 46, "swa_kv_heads": 16, "swa_head_dim": 64, "window": 2048, "arch": "swa", "max_ctx": 1048576, "config_source": "https://huggingface.co/unsloth/Muse-Spark-Open-816B-A42B/blob/main/config.json", "config_note": "official repo is gated; config read from ungated mirror" }, "moe_skip": "derived expert weights (1597.7B) exceed the model total (822.04B): the config needs a manual reading", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/meta-models/Muse-Spark-Open-816B-A42B", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" }, { "source": "ollama", "url": "https://ollama.com/library/muse-spark-open:cloud", "runs": "cloud-only", "pull": "muse-spark-open:cloud" } ] }, { "id": "glm-4.7-flash", "name": "GLM-4.7-Flash", "repo": "zai-org/GLM-4.7-Flash", "params_b": 31.22, "params_source": "https://huggingface.co/zai-org/GLM-4.7-Flash", "active_params_b": 3, "active_source": "https://huggingface.co/zai-org/GLM-4.7-Flash", "q4_file_gb": 18.31, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/unsloth/GLM-4.7-Flash-GGUF", "accessed": "2026-10-16", "license": "MIT", "context": 202752, "released": "2026-01-19", "note": "Model card: \"a 30B-A3B MoE model\".", "kv": { "layers": 47, "kv_heads": 1, "head_dim": 576, "v_dim": 0, "arch": "mla", "max_ctx": 202752, "config_source": "https://huggingface.co/zai-org/GLM-4.7-Flash/blob/main/config.json" }, "moe_skip": "derived active parameters (5.17B, 4.54B without embeddings) are more than 25% from the published 3B", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/ZhipuAI/GLM-4.7-Flash", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/glm-4.7-flash:q4_K_M", "runs": "local", "pull": "glm-4.7-flash:q4_K_M" } ] }, { "id": "glm-5.3-flash", "name": "GLM-5.3-Flash", "repo": "zai-org/GLM-5.3-Flash", "params_b": 321.32, "params_source": "https://huggingface.co/zai-org/GLM-5.3-Flash", "active_params_b": 18, "active_source": "https://huggingface.co/zai-org/GLM-5.3-Flash", "q4_file_gb": 199.71, "quant": "UD-Q4_K_XL", "q4_source": "https://huggingface.co/unsloth/GLM-5.3-Flash-GGUF", "accessed": "2026-10-16", "license": "MIT", "context": 1048576, "released": "2026-08-25", "note": "No plain Q4_K_M GGUF found; size is unsloth UD-Q4_K_XL.", "kv_skip": "config reports head_dim 0 and qk_rope_head_dim 0 for its sparse-attention layers; KV size cannot be derived", "moe_skip": "derived active parameters (25.37B, 24.10B without embeddings) are more than 25% from the published 18B", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/ZhipuAI/GLM-5.3-Flash", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/glm-5.3-flash:cloud", "runs": "cloud-only", "pull": "glm-5.3-flash:cloud" } ] }, { "id": "nemotron-3-super-120b-a12b", "name": "NVIDIA Nemotron 3 Super 120B-A12B", "repo": "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16", "params_b": 123.61, "params_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16", "active_params_b": 12, "active_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16", "q4_file_gb": 82.54, "quant": "UD-Q4_K_M", "q4_source": "https://huggingface.co/unsloth/NVIDIA-Nemotron-3-Super-120B-A12B-GGUF", "accessed": "2026-10-16", "license": "NVIDIA Open Model License", "context": 262144, "released": "2026-03-10", "status": "older", "superseded_by": "nemotron-3.5-super-vl-120b-a12b", "kv": { "layers": 8, "kv_heads": 2, "head_dim": 128, "v_dim": 128, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16/blob/main/config.json" }, "moe_skip": "derived expert weights (676.5B) exceed the model total (123.61B): the config needs a manual reading", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/nv-community/NVIDIA-Nemotron-3-Super-120B-A12B-BF16", "uploader": "other", "note": "NVIDIA community account; we found no NVIDIA page linking to it" }, { "source": "ollama", "url": "https://ollama.com/library/nemotron-3-super:120b-a12b", "runs": "local", "pull": "nemotron-3-super:120b-a12b" } ] }, { "id": "nemotron-3-nano-4b", "name": "NVIDIA Nemotron 3 Nano 4B", "repo": "nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16", "params_b": 3.97, "params_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16", "active_params_b": null, "active_source": null, "q4_file_gb": 2.9, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/unsloth/NVIDIA-Nemotron-3-Nano-4B-GGUF", "accessed": "2026-10-16", "license": "NVIDIA Open Model License", "context": 262144, "released": "2026-03-07", "kv": { "layers": 4, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/nv-community/NVIDIA-Nemotron-3-Nano-4B-BF16", "uploader": "other", "note": "NVIDIA community account; we found no NVIDIA page linking to it" }, { "source": "ollama", "url": "https://ollama.com/library/nemotron-3-nano:4b", "runs": "local", "pull": "nemotron-3-nano:4b" } ] }, { "id": "lfm2-24b-a2b", "name": "LFM2-24B-A2B", "repo": "LiquidAI/LFM2-24B-A2B", "params_b": 23.84, "params_source": "https://huggingface.co/LiquidAI/LFM2-24B-A2B", "active_params_b": 2, "active_source": "https://huggingface.co/LiquidAI/LFM2-24B-A2B", "q4_file_gb": 14.42, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/LiquidAI/LFM2-24B-A2B-GGUF", "accessed": "2026-10-16", "license": "LFM Open License 1.0", "context": 128000, "released": "2026-02-24", "status": "older", "superseded_by": "lfm2.5-24b-a2b", "kv": { "layers": 10, "kv_heads": 8, "head_dim": 64, "v_dim": 64, "arch": "hybrid", "max_ctx": 128000, "config_source": "https://huggingface.co/LiquidAI/LFM2-24B-A2B/blob/main/config.json" }, "moe_skip": "derived expert weights (24.2B) exceed the model total (23.84B): the config needs a manual reading", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/LiquidAI/LFM2-24B-A2B", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" }, { "source": "ollama", "url": "https://ollama.com/library/lfm2:24b-a2b", "runs": "local", "pull": "lfm2:24b-a2b" } ] }, { "id": "deepseek-v4-flash", "name": "DeepSeek-V4-Flash", "repo": "deepseek-ai/DeepSeek-V4-Flash", "params_b": 290.94, "params_source": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash", "active_params_b": 13, "active_source": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash", "q4_file_gb": 155.1, "quant": "UD-Q4_K_XL", "q4_source": "https://huggingface.co/unsloth/DeepSeek-V4-Flash-GGUF", "accessed": "2026-10-16", "license": "MIT", "context": 1048576, "released": "2026-04-22", "note": "Model card: 284B total, 13B activated, 1M context. No plain Q4_K_M GGUF found; size is unsloth UD-Q4_K_XL.", "kv_skip": "DeepSeek V4 compressed/indexed attention (compress_ratios, index_topk) has no verified llama.cpp KV formula yet", "moe_skip": "derived active parameters (20.41B, 19.35B without embeddings) are more than 25% from the published 13B", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/deepseek-ai/DeepSeek-V4-Flash", "uploader": "official" } ] }, { "id": "granite-4.2-30b", "name": "IBM Granite 4.2 30B", "repo": "ibm-granite/granite-4.2-30b", "params_b": 29.28, "params_source": "https://huggingface.co/ibm-granite/granite-4.2-30b", "active_params_b": null, "active_source": null, "q4_file_gb": 17.72, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/ibm-granite/granite-4.2-30b-GGUF", "accessed": "2026-10-16", "license": "Apache 2.0", "context": 131072, "released": "2026-08-25", "released_source": "https://research.ibm.com/blog/introducing-granite-4-2", "note": "Dense model. Model card: 128K native context, extendable to 512K; release date August 25, 2026.", "kv": { "layers": 64, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 131072, "config_source": "https://huggingface.co/ibm-granite/granite-4.2-30b/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/ibm-granite/granite-4.2-30b", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" }, { "source": "ollama", "url": "https://ollama.com/library/granite4.2:30b", "runs": "local", "pull": "granite4.2:30b" } ] }, { "id": "granite-4.2-8b", "name": "IBM Granite 4.2 8B", "repo": "ibm-granite/granite-4.2-8b", "params_b": 8.79, "params_source": "https://huggingface.co/ibm-granite/granite-4.2-8b", "active_params_b": null, "active_source": null, "q4_file_gb": 5.35, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/ibm-granite/granite-4.2-8b-GGUF", "accessed": "2026-10-16", "license": "Apache 2.0", "context": 131072, "released": "2026-08-25", "released_source": "https://research.ibm.com/blog/introducing-granite-4-2", "note": "Dense model. Model card: 128K native context, extendable to 512K; release date August 25, 2026.", "kv": { "layers": 40, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 131072, "config_source": "https://huggingface.co/ibm-granite/granite-4.2-8b/blob/main/config.json" }, "related": [ { "href": "/models/twil-8b/", "k": "Fine-tune", "t": "webAI TwIL-8B, a non-commercial math and code reasoning fine-tune of this model" } ], "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/ibm-granite/granite-4.2-8b", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" }, { "source": "ollama", "url": "https://ollama.com/library/granite4.2:8b", "runs": "local", "pull": "granite4.2:8b" } ] }, { "id": "granite-4.2-3b", "name": "IBM Granite 4.2 3B", "repo": "ibm-granite/granite-4.2-3b", "params_b": 3.66, "params_source": "https://huggingface.co/ibm-granite/granite-4.2-3b", "active_params_b": null, "active_source": null, "q4_file_gb": 2.24, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/ibm-granite/granite-4.2-3b-GGUF", "accessed": "2026-10-16", "license": "Apache 2.0", "context": 131072, "released": "2026-08-25", "released_source": "https://research.ibm.com/blog/introducing-granite-4-2", "note": "Dense model. Model card: 128K native context, extendable to 512K; release date August 25, 2026.", "kv": { "layers": 40, "kv_heads": 8, "head_dim": 64, "v_dim": 64, "arch": "gqa", "max_ctx": 131072, "config_source": "https://huggingface.co/ibm-granite/granite-4.2-3b/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/ibm-granite/granite-4.2-3b", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" }, { "source": "ollama", "url": "https://ollama.com/library/granite4.2:3b", "runs": "local", "pull": "granite4.2:3b" } ] }, { "id": "lfm2.5-24b-a2b", "name": "LFM2.5-24B-A2B", "repo": "LiquidAI/LFM2.5-24B-A2B", "params_b": 23.97, "params_source": "https://huggingface.co/LiquidAI/LFM2.5-24B-A2B", "active_params_b": 2, "active_source": "https://huggingface.co/LiquidAI/LFM2.5-24B-A2B", "q4_file_gb": 14.53, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/LiquidAI/LFM2.5-24B-A2B-GGUF", "accessed": "2026-10-16", "license": "LFM Open License 1.0", "context": 131072, "released": "2026-10-12", "note": "Official GGUF from Liquid AI. Reasoning model (always thinks before answering).", "kv": { "layers": 10, "kv_heads": 8, "head_dim": 64, "v_dim": 64, "arch": "hybrid", "max_ctx": 131072, "config_source": "https://huggingface.co/LiquidAI/LFM2.5-24B-A2B/blob/main/config.json" }, "moe_skip": "derived expert weights (24.2B) exceed the model total (23.97B): the config needs a manual reading", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/LiquidAI/LFM2.5-24B-A2B", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" }, { "source": "ollama", "url": "https://ollama.com/library/lfm2.5:24b-a2b-q4_K_M", "runs": "local", "pull": "lfm2.5:24b-a2b-q4_K_M" } ] }, { "id": "lfm2.5-8b-a1b", "name": "LFM2.5-8B-A1B", "repo": "LiquidAI/LFM2.5-8B-A1B", "params_b": 8.47, "params_source": "https://huggingface.co/LiquidAI/LFM2.5-8B-A1B", "active_params_b": 1.5, "active_source": "https://huggingface.co/LiquidAI/LFM2.5-8B-A1B", "q4_file_gb": 5.16, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/LiquidAI/LFM2.5-8B-A1B-GGUF", "accessed": "2026-10-16", "license": "LFM Open License 1.0", "context": 128000, "released": "2026-05-28", "note": "Model card: 8.3B total / 1.5B active. Official GGUF from Liquid AI.", "kv": { "layers": 6, "kv_heads": 8, "head_dim": 64, "v_dim": 64, "arch": "hybrid", "max_ctx": 128000, "config_source": "https://huggingface.co/LiquidAI/LFM2.5-8B-A1B/blob/main/config.json" }, "moe_skip": "derived expert weights (8.5B) exceed the model total (8.47B): the config needs a manual reading", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/LiquidAI/LFM2.5-8B-A1B", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" }, { "source": "ollama", "url": "https://ollama.com/library/lfm2.5:8b-a1b-q4_K_M", "runs": "local", "pull": "lfm2.5:8b-a1b-q4_K_M" } ] }, { "id": "nemotron-3.5-lightning-30b-a3b", "name": "NVIDIA Nemotron 3.5 Lightning 30B-A3B", "repo": "nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16", "params_b": 31.58, "params_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16", "active_params_b": 3, "active_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16", "q4_file_gb": 25.48, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF", "accessed": "2026-10-16", "license": "OpenMDW 1.1", "context": 1048576, "released": "2026-08-11", "note": "Hybrid Mamba-2 + MoE model. The card states up to 1M tokens of context; config.json defaults to 262,144. This Q4_K_M file is larger than most 30B-A3B Q4_K_M builds because bartowski keeps more of its layers at higher precision.", "kv": { "layers": 6, "kv_heads": 2, "head_dim": 128, "v_dim": 128, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16/blob/main/config.json" }, "moe_skip": "derived expert weights (99.6B) exceed the model total (31.58B): the config needs a manual reading", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/nv-community/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16", "uploader": "other", "note": "NVIDIA community account; we found no NVIDIA page linking to it" }, { "source": "ollama", "url": "https://ollama.com/library/nemotron-3.5-lightning:30b-a3b", "runs": "local", "pull": "nemotron-3.5-lightning:30b-a3b" } ] }, { "id": "nemotron-3.5-super-vl-120b-a12b", "name": "NVIDIA Nemotron 3.5 Super VL 120B-A12B", "repo": "nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16", "params_b": 124.45, "params_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16", "active_params_b": 12, "active_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16", "q4_file_gb": 70.1, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/lmstudio-community/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF", "accessed": "2026-10-16", "license": "OpenMDW 1.1", "context": 1048576, "released": "2026-10-15", "note": "Vision-language model; the parameter count includes the C-RADIO vision encoder. Image input needs the separate 1.73 GB mmproj file, which is not in the size.", "kv_skip": "llm_config omits num_hidden_layers and the hybrid layer pattern", "moe_skip": "config lacks MoE layer count", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/nv-community/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16", "uploader": "other", "note": "NVIDIA community account; we found no NVIDIA page linking to it" }, { "source": "ollama", "url": "https://ollama.com/library/nemotron-3.5-super:120b-a12b", "runs": "local", "pull": "nemotron-3.5-super:120b-a12b", "note": "Listed on Ollama as nemotron-3.5-super" } ] }, { "id": "fara1.5-4b", "name": "Microsoft Fara1.5 4B", "repo": "microsoft/Fara1.5-4B", "params_b": 4.54, "params_source": "https://huggingface.co/microsoft/Fara1.5-4B", "active_params_b": null, "active_source": null, "q4_file_gb": 2.88, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Fara1.5-4B-GGUF", "accessed": "2026-10-16", "license": "MIT", "context": 262144, "released": "2026-05-21", "note": "A computer-use agent for web browsers, not a chat model: it reads screenshots and emits click/type/scroll actions, and Microsoft recommends running it only inside its Fara or MagenticLite harness. Fine-tuned from Qwen3.5-4B; screenshots need the separate mmproj file, which is not in the size.", "kv": { "layers": 8, "kv_heads": 4, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/microsoft/Fara1.5-4B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/microsoft/Fara1.5-4B", "uploader": "other", "note": "Account labeled 'Microsoft mirror'; we found no Microsoft page linking to it" } ] }, { "id": "fara1.5-9b", "name": "Microsoft Fara1.5 9B", "repo": "microsoft/Fara1.5-9B", "params_b": 9.41, "params_source": "https://huggingface.co/microsoft/Fara1.5-9B", "active_params_b": null, "active_source": null, "q4_file_gb": 5.91, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Fara1.5-9B-GGUF", "accessed": "2026-10-16", "license": "MIT", "context": 262144, "released": "2026-05-21", "note": "A computer-use agent for web browsers, not a chat model: it reads screenshots and emits click/type/scroll actions, and Microsoft recommends running it only inside its Fara or MagenticLite harness. Fine-tuned from Qwen3.5-9B; screenshots need the separate mmproj file, which is not in the size.", "kv": { "layers": 8, "kv_heads": 4, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/microsoft/Fara1.5-9B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/microsoft/Fara1.5-9B", "uploader": "other", "note": "Account labeled 'Microsoft mirror'; we found no Microsoft page linking to it" } ] }, { "id": "fara1.5-27b", "name": "Microsoft Fara1.5 27B", "repo": "microsoft/Fara1.5-27B", "params_b": 27.36, "params_source": "https://huggingface.co/microsoft/Fara1.5-27B", "active_params_b": null, "active_source": null, "q4_file_gb": 17.53, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Fara1.5-27B-GGUF", "accessed": "2026-10-16", "license": "MIT", "context": 262144, "released": "2026-05-21", "note": "A computer-use agent for web browsers, not a chat model: it reads screenshots and emits click/type/scroll actions, and Microsoft recommends running it only inside its Fara or MagenticLite harness. Fine-tuned from Qwen3.5-27B; screenshots need the separate mmproj file, which is not in the size.", "kv": { "layers": 16, "kv_heads": 4, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/microsoft/Fara1.5-27B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/microsoft/Fara1.5-27B", "uploader": "other", "note": "Account labeled 'Microsoft mirror'; we found no Microsoft page linking to it" } ] }, { "id": "nemotron-3.5-bolt-74b-a9b", "name": "NVIDIA Nemotron 3.5 Bolt 74B-A9B", "repo": "nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16", "params_b": 74.89, "params_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16", "active_params_b": 9, "active_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16", "q4_file_gb": 50.22, "quant": "UD-Q4_K_XL", "q4_source": "https://huggingface.co/unsloth/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-GGUF", "accessed": "2026-10-18", "license": "OpenMDW 1.1", "context": 262144, "released": "2026-10-15", "released_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16", "note": "Hybrid Mamba-2 + LatentMoE vision-language model; card states 74B total, 9B active, up to 256K context. Parameter count includes the vision encoder. No plain Q4_K_M GGUF found, so the size is unsloth UD-Q4_K_XL. NVIDIA also ships a native 2-bit '24GB' GGUF (22.02 GB) built to fit 24 GB cards: https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF", "kv": { "layers": 8, "kv_heads": 2, "head_dim": 128, "v_dim": 128, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16/blob/main/config.json" }, "moe_skip": "config lacks MoE layer count, experts per token, expert width", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/nv-community/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16", "uploader": "other", "note": "NVIDIA community account; we found no NVIDIA page linking to it" }, { "source": "ollama", "url": "https://ollama.com/library/nemotron-3.5-bolt:74b-a9b", "runs": "local", "pull": "nemotron-3.5-bolt:74b-a9b" } ] }, { "id": "deepseek-v4.1-flash", "name": "DeepSeek-V4.1-Flash (1017)", "repo": "deepseek-ai/DeepSeek-V4.1-Flash-1017", "params_b": 763.21, "params_source": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017", "active_params_b": 16, "active_source": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017", "q4_file_gb": 500.71, "quant": "UD-Q4_K_XL", "q4_source": "https://huggingface.co/unsloth/DeepSeek-V4.1-Flash-1017-GGUF", "accessed": "2026-10-18", "license": "MIT", "context": 1048576, "released": "2026-09-10", "released_source": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash", "note": "Points at the Oct 17 snapshot DeepSeek-V4.1-Flash-1017, a minor update of the Sep 10 DeepSeek-V4.1-Flash with an identical config.json and parameter count. Card: 552B backbone parameters plus 196B Engram conditional memory (763B in the safetensors), 8B/16B activated (we use 16B), 1M context. No plain Q4_K_M found, so the size is unsloth UD-Q4_K_XL of the 1017 snapshot. Release date is the original V4.1-Flash release.", "kv_skip": "DeepSeek V4.1 CSA2 compressed sparse attention (compress_ratios, kv_source_layer_ids, index_topk) has no verified llama.cpp KV formula yet", "moe_skip": "no saved Hugging Face config", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/deepseek-ai/DeepSeek-V4.1-Flash-1017", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/deepseek-v4.1-flash:cloud", "runs": "cloud-only", "pull": "deepseek-v4.1-flash:cloud" } ] }, { "id": "beam-1.0-501b-a23b", "name": "Reflection Beam 1.0 501B-A23B", "repo": "reflection/Beam-1.0-501B-A23B", "params_b": 501.27, "params_source": "https://huggingface.co/reflection/Beam-1.0-501B-A23B", "active_params_b": 23, "active_source": "https://huggingface.co/reflection/Beam-1.0-501B-A23B", "q4_file_gb": 302.41, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/DevQuasar/reflection.Beam-1.0-501B-A23B-GGUF", "accessed": "2026-10-18", "license": "Apache 2.0", "context": 1048576, "released": "2026-10-07", "released_source": "https://huggingface.co/reflection/Beam-1.0-501B-A23B", "note": "Card: 501B total, 23B active, gated GQA + sliding-window attention, 1M context. Release date is the Hugging Face repo creation date.", "kv": { "layers": 9, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "swa_layers": 43, "swa_kv_heads": 8, "swa_head_dim": 128, "window": 512, "arch": "swa", "max_ctx": 1048576, "config_source": "https://huggingface.co/reflection/Beam-1.0-501B-A23B/blob/main/config.json" }, "moe_skip": "derived expert weights (523.4B) exceed the model total (501.27B): the config needs a manual reading", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "ollama", "url": "https://ollama.com/library/beam:cloud", "runs": "cloud-only", "pull": "beam:cloud" } ] }, { "id": "ternary-bonsai-2-27b", "name": "Ternary Bonsai 2 27B", "repo": "prism-ml/Ternary-Bonsai-2-27B-gguf", "params_b": 27.36, "params_source": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf", "active_params_b": null, "active_source": null, "q4_file_gb": 5.95, "quant": "PTQ1_0 ternary", "q4_source": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf", "accessed": "2026-10-18", "license": "Apache 2.0", "context": 262144, "released": "2026-09-16", "released_source": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf", "note": "Ternary (1.75 bits/weight) version of Qwen3.8-27B by PrismML; no Q4 build exists, so the size is the native PTQ1_0 GGUF (the PQ2_0 pack is 7.21 GB). Needs PrismML's llama.cpp fork (github.com/PrismML-Eng/llama.cpp): stock llama.cpp rejects these files. Release date is the Hugging Face repo creation date.", "kv": { "layers": 16, "kv_heads": 4, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/Qwen/Qwen3.8-27B/blob/main/config.json", "config_note": "GGUF-only repo has no config.json; card says the architecture is unchanged from Qwen3.8-27B, so its config is used" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/prism-ml/Ternary-Bonsai-2-27B-gguf", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" }, { "source": "ollama", "url": "https://ollama.com/library/bonsai-2:27b", "runs": "local", "pull": "bonsai-2:27b" } ] }, { "id": "k2-horizon-mova-36b-a4b", "name": "K2-Horizon-MoVA-36B-A4B", "repo": "IFM/K2-Horizon-MoVA-36B-A4B", "params_b": 37.44, "params_source": "https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B", "active_params_b": 4, "active_source": "https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B", "q4_file_gb": 22.37, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B-GGUF", "accessed": "2026-10-18", "license": "Apache 2.0", "context": 524288, "released": "2026-09-01", "released_source": "https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B", "note": "MBZUAI IFM. Card: 36B stored, 4B per token, 512K native context, Mixture-of-Values attention (MoVA). Stock llama.cpp support merged Oct 6 (PR #29535). MoVA routes the value projection through experts; the context estimate assumes llama.cpp caches K and V like standard GQA (not checked against llama.cpp memory logs). Release date is the Hugging Face repo creation date.", "kv": { "layers": 48, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 524288, "config_source": "https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B/blob/main/config.json" }, "moe_skip": "derived active parameters (13.02B, 11.74B without embeddings) are more than 25% from the published 4B", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/IFM/K2-Horizon-MoVA-36B-A4B", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" } ] }, { "id": "k2-horizon-32b", "name": "K2-Horizon-32B", "repo": "IFM/K2-Horizon-32B", "params_b": 34.78, "params_source": "https://huggingface.co/IFM/K2-Horizon-32B", "active_params_b": null, "active_source": null, "q4_file_gb": 21.08, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/IFM/K2-Horizon-32B-GGUF", "accessed": "2026-10-18", "license": "Apache 2.0", "context": 524288, "released": "2026-09-01", "released_source": "https://huggingface.co/IFM/K2-Horizon-32B", "note": "MBZUAI IFM dense model, 512K native context. Stock llama.cpp support merged Oct 6 (PR #29535). Release date is the Hugging Face repo creation date.", "kv": { "layers": 64, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 524288, "config_source": "https://huggingface.co/IFM/K2-Horizon-32B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/IFM/K2-Horizon-32B", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" } ] }, { "id": "k2-horizon-7b", "name": "K2-Horizon-7B", "repo": "IFM/K2-Horizon-7B", "params_b": 9.0, "params_source": "https://huggingface.co/IFM/K2-Horizon-7B", "active_params_b": null, "active_source": null, "q4_file_gb": 5.59, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/IFM/K2-Horizon-7B-GGUF", "accessed": "2026-10-18", "license": "Apache 2.0", "context": 524288, "released": "2026-09-01", "released_source": "https://huggingface.co/IFM/K2-Horizon-7B", "note": "MBZUAI IFM dense model (9.0B in the safetensors; the card calls it a 7B-core model), 512K native context. Stock llama.cpp support merged Oct 6 (PR #29535). Release date is the Hugging Face repo creation date.", "kv": { "layers": 36, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 524288, "config_source": "https://huggingface.co/IFM/K2-Horizon-7B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/IFM/K2-Horizon-7B", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" } ] }, { "id": "minicpm-v-4.7-35b-a3b", "name": "MiniCPM-V 4.7 35B-A3B", "repo": "openbmb/MiniCPM-V-4.7-35B-A3B", "params_b": 35.21, "params_source": "https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B", "active_params_b": 3, "active_source": "https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B", "q4_file_gb": 21.17, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B-gguf", "accessed": "2026-10-18", "license": "Apache 2.0", "context": 262144, "released": "2026-10-07", "released_source": "https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B", "note": "OpenBMB vision-language MoE on a Qwen3.5-MoE backbone; active params from the model name (A3B). Stock llama.cpp support merged Oct 10 (PR #29416). Fit uses the language-model GGUF only; the vision projector (mmproj, 1.12 GB f16) is extra when you send images, as with other vision models on this site. Release date is the Hugging Face repo creation date.", "kv": { "layers": 10, "kv_heads": 2, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B/blob/main/config.json" }, "moe": { "expert_b": 32.212, "expert_active_b": 1.007, "layers": 40, "experts": 256, "experts_per_tok": 8, "expert_width": 512, "config_source": "https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/OpenBMB/MiniCPM-V-4.7-35B-A3B", "uploader": "official" } ] }, { "id": "kolibri-1", "name": "Aleph Alpha Kolibri-1 78B-A3.5B", "repo": "Aleph-Alpha/Kolibri-1", "params_b": 78.1, "params_source": "https://huggingface.co/Aleph-Alpha/Kolibri-1", "active_params_b": 3.46, "active_source": "https://huggingface.co/Aleph-Alpha/Kolibri-1", "q4_file_gb": 49.76, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Kolibri-1-GGUF", "accessed": "2026-10-18", "license": "Apache 2.0", "context": 1048576, "released": "2026-10-02", "released_source": "https://huggingface.co/Aleph-Alpha/Kolibri-1", "note": "Card: 78.1B total, 3.46B active, 1M context (262K recommended; config.json defaults to 262,144). Stock llama.cpp support merged Oct 17 (PR #30673), so older llama.cpp builds can't load it. No official GGUF: the size is bartowski's community Q4_K_M (Hob-forge's earlier Q4_K_M is 47.45 GB). Release date is the Hugging Face repo creation date.", "kv": { "layers": 10, "kv_heads": 4, "head_dim": 128, "v_dim": 128, "swa_layers": 40, "swa_kv_heads": 4, "swa_head_dim": 128, "window": 513, "arch": "swa", "max_ctx": 262144, "config_source": "https://huggingface.co/Aleph-Alpha/Kolibri-1/blob/main/config.json" }, "moe": { "expert_b": 75.497, "expert_active_b": 1.18, "layers": 50, "experts": 384, "experts_per_tok": 6, "expert_width": 512, "config_source": "https://huggingface.co/Aleph-Alpha/Kolibri-1/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Aleph-Alpha/Kolibri-1", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" } ] }, { "id": "ternary-bonsai-2-9b", "name": "Ternary Bonsai 2 9B", "repo": "prism-ml/Ternary-Bonsai-2-9B-gguf", "params_b": 8.95, "params_source": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-9B-gguf", "active_params_b": null, "active_source": null, "q4_file_gb": 2.0, "quant": "PTQ1_0 ternary", "q4_source": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-9B-gguf", "accessed": "2026-10-18", "license": "Apache 2.0", "context": 262144, "released": "2026-09-23", "released_source": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-9B-gguf", "note": "Ternary version of Qwen3.5-9B by PrismML; card: ~8.95B language model plus a ~0.46B vision tower. No Q4 build exists, so the size is the native PTQ1_0 GGUF (PQ2_0 is 2.42 GB). Needs PrismML's llama.cpp fork (github.com/PrismML-Eng/llama.cpp) for the PTQ1_0 file; stock llama.cpp rejects PTQ1_0 and PQ2_0 (support PRs #29672/#29676 still open). The card says upstream llama.cpp runs only the Q2_0 pack (2.56 GB here). Release date is the Hugging Face repo creation date; PrismML's public launch was Oct 15.", "kv": { "layers": 8, "kv_heads": 4, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/Qwen/Qwen3.5-9B/blob/main/config.json", "config_note": "GGUF-only repo has no config.json; card says the architecture is unchanged from Qwen3.5-9B, so its config is used" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/prism-ml/Ternary-Bonsai-2-9B-gguf", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" }, { "source": "ollama", "url": "https://ollama.com/library/bonsai-2:9b", "runs": "local", "pull": "bonsai-2:9b" } ] }, { "id": "ternary-bonsai-2-4b", "name": "Ternary Bonsai 2 4B", "repo": "prism-ml/Ternary-Bonsai-2-4B-gguf", "params_b": 4.21, "params_source": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-4B-gguf", "active_params_b": null, "active_source": null, "q4_file_gb": 0.95, "quant": "PTQ1_0 ternary", "q4_source": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-4B-gguf", "accessed": "2026-10-18", "license": "Apache 2.0", "context": 262144, "released": "2026-09-23", "released_source": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-4B-gguf", "note": "Ternary version of Qwen3.5-4B by PrismML; card: ~4.21B language model plus a ~0.33B vision tower. No Q4 build exists, so the size is the native PTQ1_0 GGUF (PQ2_0 is 1.15 GB). Needs PrismML's llama.cpp fork (github.com/PrismML-Eng/llama.cpp) for the PTQ1_0 file; stock llama.cpp rejects PTQ1_0 and PQ2_0 (support PRs #29672/#29676 still open). The card says upstream llama.cpp runs only the Q2_0 pack (1.21 GB here). Release date is the Hugging Face repo creation date; PrismML's public launch was Oct 15.", "kv": { "layers": 8, "kv_heads": 4, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/Qwen/Qwen3.5-4B/blob/main/config.json", "config_note": "GGUF-only repo has no config.json; card says the architecture is unchanged from Qwen3.5-4B, so its config is used" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/prism-ml/Ternary-Bonsai-2-4B-gguf", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" }, { "source": "ollama", "url": "https://ollama.com/library/bonsai-2:4b", "runs": "local", "pull": "bonsai-2:4b" } ] }, { "id": "ternary-bonsai-2-2b", "name": "Ternary Bonsai 2 2B", "repo": "prism-ml/Ternary-Bonsai-2-2B-gguf", "params_b": 1.88, "params_source": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-2B-gguf", "active_params_b": null, "active_source": null, "q4_file_gb": 0.43, "quant": "PTQ1_0 ternary", "q4_source": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-2B-gguf", "accessed": "2026-10-18", "license": "Apache 2.0", "context": 262144, "released": "2026-09-23", "released_source": "https://huggingface.co/prism-ml/Ternary-Bonsai-2-2B-gguf", "note": "Ternary version of Qwen3.5-2B by PrismML; card: ~1.88B language model plus a ~0.33B vision tower. No Q4 build exists, so the size is the native PTQ1_0 GGUF (PQ2_0 is 0.52 GB). Needs PrismML's llama.cpp fork (github.com/PrismML-Eng/llama.cpp) for the PTQ1_0 file; stock llama.cpp rejects PTQ1_0 and PQ2_0 (support PRs #29672/#29676 still open). The card says upstream llama.cpp runs only the Q2_0 pack (0.55 GB here). Release date is the Hugging Face repo creation date; PrismML's public launch was Oct 15.", "kv": { "layers": 6, "kv_heads": 2, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/Qwen/Qwen3.5-2B/blob/main/config.json", "config_note": "GGUF-only repo has no config.json; card says the architecture is unchanged from Qwen3.5-2B, so its config is used" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/prism-ml/Ternary-Bonsai-2-2B-gguf", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" }, { "source": "ollama", "url": "https://ollama.com/library/bonsai-2:2b", "runs": "local", "pull": "bonsai-2:2b" } ] }, { "id": "nemotron-3.5-bolt-74b-a9b-24gb", "name": "NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB (2-bit)", "repo": "nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB", "params_b": 74.89, "params_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB", "active_params_b": 9, "active_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB", "q4_file_gb": 22.02, "quant": "native 2-bit", "q4_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF", "accessed": "2026-10-18", "license": "OpenMDW 1.1", "context": 262144, "released": "2026-10-07", "released_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF", "note": "NVIDIA's official 24GB build of Nemotron 3.5 Bolt: native 2-bit quantization, not Q4, so expect lower quality than the 50.22 GB UD-Q4_K_XL row. Card: made to fit 24 GB of memory; the GGUF is NVIDIA's own for llama.cpp. Parameter count is the full BF16 model's. Fit ignores the 1.73 GB vision projector (mmproj). Release date is the Hugging Face repo creation date.", "kv": { "layers": 8, "kv_heads": 2, "head_dim": 128, "v_dim": 128, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB/blob/main/config.json" }, "moe_skip": "config lacks MoE layer count, experts per token, expert width", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/nv-community/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB", "uploader": "other", "note": "NVIDIA community account; we found no NVIDIA page linking to it" } ] }, { "id": "mimo-v2.6-flash", "name": "Xiaomi MiMo-V2.6-Flash 309B-A15B", "repo": "XiaomiMiMo/MiMo-V2.6-Flash-RL", "params_b": 310.76, "params_source": "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL", "active_params_b": 15, "active_source": "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL", "q4_file_gb": 167.37, "quant": "MXFP4", "q4_source": "https://huggingface.co/ggml-org/MiMo-V2.6-Flash-RL-GGUF", "accessed": "2026-10-18", "license": "MIT", "context": 1048576, "released": "2026-09-21", "released_source": "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL", "note": "Card: 309B total, 15B activated, 1M context. RL release (the MOPD release of Sep 27 has the same size). Runs in stock llama.cpp (mimo2 architecture; MiMo-V2.6 conversion merged Sep 22, PR #29257). No Q4_K_M found; the size is ggml-org's official MXFP4 GGUF. Release date is the Hugging Face repo creation date.", "kv": { "layers": 9, "kv_heads": 4, "head_dim": 192, "v_dim": 128, "swa_layers": 39, "swa_kv_heads": 8, "swa_head_dim": 160, "window": 128, "arch": "swa", "max_ctx": 1048576, "config_source": "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL/blob/main/config.json", "config_note": "hybrid_layer_pattern: 9 full + 39 sliding-window (128) layers; SWA K/V dims 192/128 averaged to 160 so the byte count is exact" }, "moe": { "expert_b": 302.795, "expert_active_b": 9.462, "layers": 47, "experts": 256, "experts_per_tok": 8, "expert_width": 2048, "config_source": "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/XiaomiMiMo/MiMo-V2.6-Flash-RL", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/mimo-v2.6-flash:cloud", "runs": "cloud-only", "pull": "mimo-v2.6-flash:cloud" } ] }, { "id": "mimo-v2.6-distill-qwen-9b", "name": "Xiaomi MiMo-V2.6-Distill-Qwen-9B", "repo": "XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B", "params_b": 9.41, "params_source": "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B", "active_params_b": null, "active_source": null, "q4_file_gb": 5.84, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/MiMo-V2.6-Distill-Qwen-9B-GGUF", "accessed": "2026-10-18", "license": "MIT", "context": 262144, "released": "2026-09-21", "released_source": "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B", "note": "Fine-tune of Qwen3.5-9B distilled from MiMo-V2.6; runs in stock llama.cpp (qwen35 architecture). Fit ignores the 0.92 GB vision projector. Release date is the Hugging Face repo creation date.", "kv": { "layers": 8, "kv_heads": 4, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/mimo-v2.6:9b", "runs": "local", "pull": "mimo-v2.6:9b", "note": "Listed on Ollama as mimo-v2.6" } ] }, { "id": "mellum2.1-12b-a2.5b", "name": "JetBrains Mellum2.1 12B-A2.5B Thinking", "repo": "JetBrains/Mellum2.1-12B-A2.5B-Thinking", "params_b": 12.15, "params_source": "https://huggingface.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking", "active_params_b": 2.5, "active_source": "https://huggingface.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking", "q4_file_gb": 8.07, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF", "accessed": "2026-10-18", "license": "Apache 2.0", "context": 131072, "released": "2026-09-20", "released_source": "https://huggingface.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking", "note": "Card: 12B total, 2.5B active, 131K context; coding-focused MoE. The mellum architecture is in stock llama.cpp. Release date is the Hugging Face repo creation date.", "kv": { "layers": 7, "kv_heads": 4, "head_dim": 128, "v_dim": 128, "swa_layers": 21, "swa_kv_heads": 4, "swa_head_dim": 128, "window": 1024, "arch": "swa", "max_ctx": 131072, "config_source": "https://huggingface.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking/blob/main/config.json" }, "moe": { "expert_b": 11.098, "expert_active_b": 1.387, "layers": 28, "experts": 64, "experts_per_tok": 8, "expert_width": 896, "config_source": "https://huggingface.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/JetBrains/Mellum2.1-12B-A2.5B-Thinking", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" } ] }, { "id": "mistral-small-4-119b", "name": "Mistral Small 4 119B", "repo": "mistralai/Mistral-Small-4-119B-2603", "params_b": 119.4, "params_source": "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603", "active_params_b": 6.5, "active_source": "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603", "q4_file_gb": 72.64, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/mistralai_Mistral-Small-4-119B-2603-GGUF", "accessed": "2026-10-18", "license": "Apache 2.0", "context": 262144, "released": "2026-01-23", "released_source": "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603", "note": "Card: 119B total, 6.5B active, 256K context, MLA attention. The mistral4 architecture is in stock llama.cpp. Fit ignores the 0.87 GB vision projector. Release date is the Hugging Face repo creation date.", "kv": { "layers": 36, "kv_heads": 1, "head_dim": 320, "v_dim": 0, "arch": "mla", "max_ctx": 262144, "config_source": "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603/blob/main/config.json" }, "moe": { "expert_b": 115.964, "expert_active_b": 3.624, "layers": 36, "experts": 128, "experts_per_tok": 4, "expert_width": 2048, "config_source": "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/mistralai/Mistral-Small-4-119B-2603", "uploader": "other", "note": "Account named Mistral AI; we found no Mistral page linking to it" } ] }, { "id": "mistral-medium-3.5-128b", "name": "Mistral Medium 3.5 128B", "repo": "mistralai/Mistral-Medium-3.5-128B", "params_b": 127.7, "params_source": "https://huggingface.co/mistralai/Mistral-Medium-3.5-128B", "active_params_b": null, "active_source": null, "q4_file_gb": 78.41, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/mistralai_Mistral-Medium-3.5-128B-GGUF", "accessed": "2026-10-18", "license": "Modified MIT (revenue limit)", "context": 262144, "released": "2026-03-31", "released_source": "https://huggingface.co/mistralai/Mistral-Medium-3.5-128B", "note": "Card: dense 128B, 256K context. License is Mistral's Modified MIT with exceptions for large-revenue companies. Runs in stock llama.cpp (mistral3). Fit ignores the 5.38 GB vision projector. Release date is the Hugging Face repo creation date.", "kv": { "layers": 88, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 262144, "config_source": "https://huggingface.co/mistralai/Mistral-Medium-3.5-128B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/mistralai/Mistral-Medium-3.5-128B", "uploader": "other", "note": "Account named Mistral AI; we found no Mistral page linking to it" }, { "source": "ollama", "url": "https://ollama.com/library/mistral-medium-3.5:128b", "runs": "local", "pull": "mistral-medium-3.5:128b" } ] }, { "id": "qwen3.5-122b-a10b", "name": "Qwen3.5-122B-A10B", "repo": "Qwen/Qwen3.5-122B-A10B", "params_b": 125.09, "params_source": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B", "active_params_b": 10, "active_source": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B", "q4_file_gb": 77.62, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Qwen_Qwen3.5-122B-A10B-GGUF", "accessed": "2026-10-18", "license": "Apache 2.0", "context": 262144, "released": "2026-02-24", "released_source": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B", "note": "Card: 122B total, 10B activated, 262K native context. Runs in stock llama.cpp (qwen35moe). Fit ignores the 0.91 GB vision projector. Release date is the Hugging Face repo creation date.", "kv": { "layers": 12, "kv_heads": 2, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B/blob/main/config.json" }, "moe": { "expert_b": 115.964, "expert_active_b": 3.624, "layers": 48, "experts": 256, "experts_per_tok": 8, "expert_width": 1024, "config_source": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Qwen/Qwen3.5-122B-A10B", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/qwen3.5:122b-a10b", "runs": "local", "pull": "qwen3.5:122b-a10b" } ] }, { "id": "step-3.7-flash", "name": "StepFun Step 3.7 Flash", "repo": "stepfun-ai/Step-3.7-Flash", "params_b": 201.37, "params_source": "https://huggingface.co/stepfun-ai/Step-3.7-Flash", "active_params_b": 11, "active_source": "https://huggingface.co/stepfun-ai/Step-3.7-Flash", "q4_file_gb": 122.09, "quant": "UD-Q4_K_M", "q4_source": "https://huggingface.co/unsloth/Step-3.7-Flash-GGUF", "accessed": "2026-10-18", "license": "Apache 2.0", "context": 262144, "released": "2026-05-23", "released_source": "https://huggingface.co/stepfun-ai/Step-3.7-Flash", "note": "Card: 198B (196B language + 1.8B vision), about 11B active. Stock llama.cpp support merged Jun 2 (PR #23845). No plain Q4_K_M found (StepFun's own GGUF has Q4_K_S 111.50 GB); the size is unsloth UD-Q4_K_M. Fit ignores the vision projector. Release date is the Hugging Face repo creation date.", "kv": { "layers": 12, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "swa_layers": 33, "swa_kv_heads": 8, "swa_head_dim": 128, "window": 512, "arch": "swa", "max_ctx": 262144, "config_source": "https://huggingface.co/stepfun-ai/Step-3.7-Flash/blob/main/config.json", "config_note": "KV heads from num_attention_groups (8); the 3 extra layer_types entries are MTP layers, not counted" }, "moe_skip": "config lacks expert count, experts per token", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/stepfun-ai/Step-3.7-Flash", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" } ] }, { "id": "inkling-small", "name": "Thinking Machines Inkling-Small", "repo": "thinkingmachines/Inkling-Small", "params_b": 265.96, "params_source": "https://huggingface.co/thinkingmachines/Inkling-Small", "active_params_b": 12, "active_source": "https://huggingface.co/thinkingmachines/Inkling-Small", "q4_file_gb": 159.34, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/DevQuasar/thinkingmachines.Inkling-Small-GGUF", "accessed": "2026-10-18", "license": "Apache 2.0", "context": 1048576, "released": "2026-07-27", "released_source": "https://huggingface.co/thinkingmachines/Inkling-Small", "status": "older", "superseded_by": "inkling-2-small", "related": [ { "href": "/models/inkling-2-small/", "k": "Model", "t": "Inkling 2 Small, the newer model in Thinking Machines' Small line" } ], "note": "Card: 276B total, 12B active, local + global hybrid attention. NOT supported by stock llama.cpp: the Inkling architecture PR (#25731) is still open, so these GGUFs need a patched build. Context is config.json's model_max_length (the card states none). Release date is the Hugging Face repo creation date.", "kv": { "layers": 7, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "swa_layers": 35, "swa_kv_heads": 8, "swa_head_dim": 128, "window": 512, "arch": "swa", "max_ctx": 1048576, "config_source": "https://huggingface.co/thinkingmachines/Inkling-Small/blob/main/config.json", "config_note": "local_layer_ids: 35 sliding-window (512) + 7 global layers; Inkling relative-position extras (d_rel) ignored" }, "moe_skip": "config lacks expert width", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/thinkingmachines/Inkling-Small", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" } ] }, { "id": "ling-3.1-flash", "name": "inclusionAI Ling-3.1-flash 560B-A25B", "repo": "inclusionAI/Ling-3.1-flash", "params_b": 572.12, "params_source": "https://huggingface.co/inclusionAI/Ling-3.1-flash", "active_params_b": 25, "active_source": "https://huggingface.co/inclusionAI/Ling-3.1-flash", "q4_file_gb": 357.97, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Ling-3.1-flash-GGUF", "accessed": "2026-10-19", "license": "MIT", "context": 1048576, "released": "2026-10-13", "released_source": "https://huggingface.co/inclusionAI/Ling-3.1-flash", "note": "Card: 560B total, 25B active, hybrid-linear MoE (63 KDA linear-attention + 9 gated MLA layers), 1M-token context. Parameter count is the safetensors total of the checkpoint (bartowski's GGUF card also says 572B), which includes the MTP layer. config.json's max_position_embeddings is 262,144, below the card's 1M. The BailingMoeV3 architecture is in stock llama.cpp (bartowski quantized with release b11627). Release date is the Hugging Face repo creation date.", "kv": { "layers": 9, "kv_heads": 1, "head_dim": 576, "v_dim": 0, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/inclusionAI/Ling-3.1-flash/blob/main/config.json", "config_note": "72 layers in groups of 8 (layer_group_size): every 8th layer is gated MLA, the other 63 are KDA linear attention with a fixed-size state (arch hybrid, so the speed estimate adds the recurrent-state cost); each MLA layer caches one latent of kv_lora_rank 512 + qk_rope_head_dim 64 (llama.cpp convert: conversion/bailingmoe3.py)" }, "moe_skip": "derived active parameters (33.07B, 31.78B without embeddings) are more than 25% from the published 25B", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/inclusionAI/Ling-3.1-flash", "uploader": "official" } ] }, { "id": "minicpm-v-4.7-1b", "name": "MiniCPM-V 4.7 1B", "repo": "openbmb/MiniCPM-V-4.7-1B", "params_b": 1.55, "params_source": "https://huggingface.co/openbmb/MiniCPM-V-4.7-1B", "q4_file_gb": 0.67, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/openbmb/MiniCPM-V-4.7-1B-gguf", "accessed": "2026-10-19", "license": "Apache 2.0", "context": 262144, "released": "2026-10-06", "released_source": "https://huggingface.co/openbmb/MiniCPM-V-4.7-1B", "note": "OpenBMB on-device vision-language model (dense 1B language backbone on Qwen3.5 text, LLaVA-UHD v4 early-compression vision). Card positions it as the phone/tablet sibling of MiniCPM-V 4.7 35B-A3B. Fit uses the language-model GGUF only; the vision projector is extra when you send images. Release date is the Hugging Face repo creation date.", "kv": { "layers": 6, "kv_heads": 2, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/openbmb/MiniCPM-V-4.7-1B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/OpenBMB/MiniCPM-V-4.7-1B", "uploader": "official" } ] }, { "id": "flint-40b-a2.5b", "name": "Recursive Flint 40B-A2.5B", "repo": "RecursiveSI/flint", "params_b": 39.8, "params_source": "https://huggingface.co/RecursiveSI/flint", "active_params_b": 2.5, "active_source": "https://huggingface.co/RecursiveSI/flint", "q4_file_gb": 25.99, "quant": "Q4_1", "q4_source": "https://huggingface.co/RecursiveSI/flint-int4-gguf", "accessed": "2026-10-28", "license": "Apache 2.0", "context": 262144, "released": "2026-10-21", "released_source": "https://huggingface.co/RecursiveSI/flint", "note": "Card: 39.8B total, ~2.5B active, 262K documented context; coding-agent MoE. Official GGUF is Q4_1 (26.0 GB, F32 embeddings, larger than a typical Q4_K_M). Flint architecture merged into llama.cpp (ggml-org/llama.cpp PR #30928). Release date is the Hugging Face repo creation date.", "quants": [ { "quant": "Q4_K_M", "bytes": 23387911136, "source": "https://huggingface.co/ggml-org/Flint-GGUF", "by": "ggml-org", "file": "Flint-Q4_K_M.gguf" }, { "quant": "Q4_K_M", "bytes": 24166841184, "source": "https://huggingface.co/lmstudio-community/flint-GGUF", "by": "lmstudio-community", "file": "flint-Q4_K_M.gguf" }, { "quant": "Q4_K_M", "bytes": 25346651360, "source": "https://huggingface.co/bartowski/RecursiveSI_flint-GGUF", "by": "bartowski", "file": "RecursiveSI_flint-Q4_K_M.gguf" }, { "quant": "Q4_1", "bytes": 25992966976, "source": "https://huggingface.co/RecursiveSI/flint-int4-gguf", "by": "RecursiveSI", "file": "flint-int4-q4_1.gguf" } ], "kv": { "layers": 14, "kv_heads": 4, "head_dim": 128, "v_dim": 128, "swa_layers": 35, "swa_kv_heads": 4, "swa_head_dim": 128, "window": 513, "arch": "swa", "max_ctx": 262144, "config_source": "https://huggingface.co/RecursiveSI/flint/blob/main/config.json" }, "moe": { "expert_b": 37.849, "expert_active_b": 0.887, "layers": 47, "experts": 256, "experts_per_tok": 6, "expert_width": 512, "config_source": "https://huggingface.co/RecursiveSI/flint/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/RecursiveSI/flint", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" }, { "source": "ollama", "url": "https://ollama.com/library/flint:40b-a2b", "runs": "local", "pull": "flint:40b-a2b" } ] }, { "id": "spark-x2.6-4b", "name": "Spark-X2.6-4B", "repo": "XHToken/Spark-X2.6-4B", "params_b": 4.11, "params_source": "https://huggingface.co/XHToken/Spark-X2.6-4B", "active_params_b": null, "active_source": null, "q4_file_gb": 3.54, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/XHToken/Spark-X2.6-4B-GGUF", "accessed": "2026-10-22", "license": "Apache 2.0", "context": 1048576, "released": "2026-10-19", "released_source": "https://huggingface.co/XHToken/Spark-X2.6-4B", "note": "iFlytek SparkLLM (XHToken) on-device 4B dense model; same architecture as Spark-X2.5-4B with improved post-training. Card documents a 1,048,576-token max context. Official Q4_K_M GGUF is 3.54 GB. spark2_5 architecture supported in mainline llama.cpp since b10828 (ggml-org/llama.cpp PR #27868). Release date is the Hugging Face repo creation date.", "kv": { "layers": 9, "kv_heads": 4, "head_dim": 256, "v_dim": 256, "swa_layers": 27, "swa_kv_heads": 4, "swa_head_dim": 256, "window": 512, "arch": "swa", "max_ctx": 1048576, "config_source": "https://huggingface.co/XHToken/Spark-X2.6-4B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/XHToken/Spark-X2.6-4B", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" } ] }, { "id": "minimax-m3.1-flash", "name": "MiniMax-M3.1-Flash", "repo": "MiniMaxAI/MiniMax-M3.1-Flash", "params_b": 427.05, "params_source": "https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash", "active_params_b": 23, "active_source": "https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash", "q4_file_gb": 258.01, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/AtomicChat/MiniMax-M3.1-Flash-GGUF", "accessed": "2026-10-23", "license": "MiniMax Community License", "context": 1048576, "released": "2026-10-17", "released_source": "https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash", "note": "Card: approximately 428B total, 23B activated per token, text/image/video in, 1M context, MiniMax Sparse Attention on every layer. Parameter count is the checkpoint's tensor shapes with NVFP4 expert tensors (weight_packed, U8) counted at 2 parameters per byte and quant-scale tensors excluded: 427.05B. Hugging Face's safetensors total is 220.49B stored elements, because those experts are packed. The weights ship already quantized (NVFP4 routed experts, MXFP8 attention) plus a separate DSpark draft that is not in the GGUF. Q4_K_M is AtomicChat's plain quant for stock llama.cpp (minimax-m3, no patch): 258.01 GB, the sum of the six Q4_K_M shards. The same repo has Q3_K_M at 203.49 GB and Q2_K at 155.91 GB (its own page: minimax-m3.1-flash-q2-k). Fit ignores the 1.73 GB vision projector. Commercial use of the MiniMax Community License needs attribution and, above $20M yearly revenue, written authorization. Release date is the Hugging Face repo creation date. Announced at https://x.com/MiniMax_AI/status/2113436864904610008.", "kv": { "layers": 60, "kv_heads": 4, "head_dim": 128, "v_dim": 128, "idx_layers": 57, "idx_kv_heads": 1, "idx_head_dim": 128, "idx_v_dim": 128, "arch": "gqa", "max_ctx": 1048576, "config_source": "https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash/blob/main/config.json", "config_note": "sparse_attention_freq is 1 on all 60 layers, including the 3 leading dense-MLP layers (moe_layer_freq). llama.cpp minimax-m3 sets leading_dense_block_count from those leading zeros and allocates the indexer cache only on the other 57 layers: 1 KV head, K and V both sparse_index_dim 128, same element type as the main cache (src/llama-kv-cache-msa.cpp). The graph writes indexer K only; V is still allocated" }, "moe_skip": "config lacks expert width", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/MiniMax/MiniMax-M3.1-Flash", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/minimax-m3.1-flash:cloud", "runs": "cloud-only", "pull": "minimax-m3.1-flash:cloud" } ] }, { "id": "minimax-m3.1-flash-q2-k", "name": "MiniMax-M3.1-Flash Q2_K", "repo": "MiniMaxAI/MiniMax-M3.1-Flash", "params_b": 427.05, "params_source": "https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash", "active_params_b": 23, "active_source": "https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash", "q4_file_gb": 155.91, "quant": "Q2_K", "q4_source": "https://huggingface.co/AtomicChat/MiniMax-M3.1-Flash-GGUF", "accessed": "2026-10-23", "license": "MiniMax Community License", "context": 1048576, "released": "2026-10-17", "released_source": "https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash", "note": "Same checkpoint as MiniMax-M3.1-Flash. This row is AtomicChat's plain Q2_K (155.91 GB, four shards in the same repo), not Q4_K_M. Their card says a 2-bit file without an importance matrix loses noticeably more quality than the larger quants; an importance-matrix build was not in the repo on 2026-10-23. Q3_K_M in that repo is 203.49 GB. Fit ignores the 1.73 GB vision projector. Use the Q4_K_M page (minimax-m3.1-flash, 258.01 GB) when the machine has room for it.", "kv": { "layers": 60, "kv_heads": 4, "head_dim": 128, "v_dim": 128, "idx_layers": 57, "idx_kv_heads": 1, "idx_head_dim": 128, "idx_v_dim": 128, "arch": "gqa", "max_ctx": 1048576, "config_source": "https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash/blob/main/config.json", "config_note": "sparse_attention_freq is 1 on all 60 layers, including the 3 leading dense-MLP layers (moe_layer_freq). llama.cpp minimax-m3 sets leading_dense_block_count from those leading zeros and allocates the indexer cache only on the other 57 layers: 1 KV head, K and V both sparse_index_dim 128, same element type as the main cache (src/llama-kv-cache-msa.cpp). The graph writes indexer K only; V is still allocated" }, "moe_skip": "config lacks expert width", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/MiniMax/MiniMax-M3.1-Flash", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/minimax-m3.1-flash:cloud", "runs": "cloud-only", "pull": "minimax-m3.1-flash:cloud" } ] }, { "id": "minimax-m3.1-flash-unsloth-ud-q2-k-xl", "name": "MiniMax-M3.1-Flash Unsloth 2-bit", "repo": "MiniMaxAI/MiniMax-M3.1-Flash", "params_b": 427.05, "params_source": "https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash", "active_params_b": 23, "active_source": "https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash", "q4_file_gb": 143.91, "quant": "UD-Q2_K_XL", "q4_source": "https://huggingface.co/unsloth/MiniMax-M3.1-Flash-GGUF/tree/main/UD-Q2_K_XL", "accessed": "2026-10-26", "license": "MiniMax Community License", "context": 1048576, "released": "2026-10-17", "released_source": "https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash", "note": "Same checkpoint as MiniMax-M3.1-Flash. This row is Unsloth's UD-Q2_K_XL (Unsloth Dynamic 2-bit): 143,910,068,580 bytes = 143.91 GB, the sum of the four shards in the UD-Q2_K_XL folder. Unsloth says this 2-bit quant reaches 91.5% top-1 accuracy and is about 45% smaller; that is Unsloth's own claim, not verified by RigPerDollar. Unsloth recommends budgeting 160 GB+ of RAM + VRAM (or unified memory) at 16K context; its table lists about 150 GB for 2-bit. Needs a llama.cpp build with MiniMax-M3.1 support (llama.cpp PR #31192) or Unsloth Desktop. Fit ignores the vision projector. Guide: https://unsloth.ai/docs/models/minimax-m3.1-flash . Announced at https://x.com/UnslothAI/status/2114730104228286936. Use the Q4_K_M page (minimax-m3.1-flash, 258.01 GB) when the machine has room for it; AtomicChat's plain Q2_K is minimax-m3.1-flash-q2-k (155.91 GB).", "kv": { "layers": 60, "kv_heads": 4, "head_dim": 128, "v_dim": 128, "idx_layers": 57, "idx_kv_heads": 1, "idx_head_dim": 128, "idx_v_dim": 128, "arch": "gqa", "max_ctx": 1048576, "config_source": "https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash/blob/main/config.json", "config_note": "sparse_attention_freq is 1 on all 60 layers, including the 3 leading dense-MLP layers (moe_layer_freq). llama.cpp minimax-m3 sets leading_dense_block_count from those leading zeros and allocates the indexer cache only on the other 57 layers: 1 KV head, K and V both sparse_index_dim 128, same element type as the main cache (src/llama-kv-cache-msa.cpp). The graph writes indexer K only; V is still allocated" }, "moe_skip": "config lacks expert width", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/MiniMax/MiniMax-M3.1-Flash", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/minimax-m3.1-flash:cloud", "runs": "cloud-only", "pull": "minimax-m3.1-flash:cloud" } ] }, { "id": "nemotron-3.5-bolt-e4b", "name": "NVIDIA Nemotron 3.5 Bolt E4B", "repo": "nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-BF16", "params_b": 8.0, "params_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-BF16", "active_params_b": null, "active_source": null, "q4_file_gb": 6.18, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB", "accessed": "2026-10-23", "license": "OpenMDW 1.1", "context": 65536, "released": "2026-10-05", "released_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-BF16", "note": "Text, image and audio in. Built on Gemma 4 E4B (Gemma4ForConditionalGeneration): \"E4B\" means about 4B effective parameters, and the 8.0B count is Hugging Face's safetensors total of 7,996,156,490, which includes the vision and audio encoders. Q4_K_M is NVIDIA's official GGUF from their \"4GB\" repo (Nemotron-3.5-Bolt-E4B-Q4_K_M.gguf, 6,180,446,208 bytes), so despite the repo name the file is 6.18 GB. Unsloth's UD quants are smaller; every size is in the table below. NVIDIA pitches the 4GB build for local assistants and tool use (https://x.com/i/status/2113657112408785153); Unsloth says it runs on 6 GB of RAM (https://x.com/i/status/2113658964365971681). Those are their claims, not checked here. Fit ignores the vision projector (mmproj-F16 990,373,568 bytes, mmproj-Q8_0 559,875,776 bytes), which is extra if you send images. Release date is the Hugging Face repo creation date of the BF16 weights.", "quants": [ { "quant": "UD-IQ2_M", "bytes": 3531578560, "source": "https://huggingface.co/unsloth/NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF", "by": "Unsloth" }, { "quant": "UD-IQ3_XXS", "bytes": 3703968960, "source": "https://huggingface.co/unsloth/NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF", "by": "Unsloth" }, { "quant": "UD-Q2_K_XL", "bytes": 3742102720, "source": "https://huggingface.co/unsloth/NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF", "by": "Unsloth" }, { "quant": "UD-Q3_K_XL", "bytes": 4219133120, "source": "https://huggingface.co/unsloth/NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF", "by": "Unsloth" }, { "quant": "UD-Q4_K_XL", "bytes": 4758125760, "source": "https://huggingface.co/unsloth/NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF", "by": "Unsloth" }, { "quant": "UD-Q5_K_XL", "bytes": 5832930496, "source": "https://huggingface.co/unsloth/NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF", "by": "Unsloth" }, { "quant": "Q4_K_M", "bytes": 6180446208, "source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB", "by": "NVIDIA", "file": "Nemotron-3.5-Bolt-E4B-Q4_K_M.gguf" }, { "quant": "UD-Q6_K_XL", "bytes": 7258879168, "source": "https://huggingface.co/unsloth/NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF", "by": "Unsloth" }, { "quant": "UD-Q8_K_XL", "bytes": 8192955584, "source": "https://huggingface.co/unsloth/NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF", "by": "Unsloth" }, { "quant": "BF16", "bytes": 15053099936, "source": "https://huggingface.co/unsloth/NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF", "by": "Unsloth" }, { "quant": "mmproj-Q8_0", "bytes": 559875776, "source": "https://huggingface.co/unsloth/NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF", "by": "Unsloth", "projector": true }, { "quant": "mmproj-F16", "bytes": 990373568, "source": "https://huggingface.co/unsloth/NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF", "by": "Unsloth", "projector": true } ], "kv": { "layers": 4, "kv_heads": 2, "head_dim": 512, "v_dim": 512, "swa_layers": 20, "swa_kv_heads": 2, "swa_head_dim": 256, "window": 512, "shared_kv_layers": 18, "arch": "swa", "max_ctx": 65536, "config_source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-BF16/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/nv-community/NVIDIA-Nemotron-3.5-Bolt-E4B-BF16", "uploader": "other", "note": "NVIDIA community account; we found no NVIDIA page linking to it" }, { "source": "ollama", "url": "https://ollama.com/library/nemotron-3.5-bolt:e4b", "runs": "local", "pull": "nemotron-3.5-bolt:e4b" } ] }, { "id": "nanbeige4.3-33b", "name": "Nanbeige4.3-33B", "repo": "Nanbeige/Nanbeige4.3-33B", "params_b": 33.12, "params_source": "https://huggingface.co/Nanbeige/Nanbeige4.3-33B", "active_params_b": null, "active_source": null, "q4_file_gb": 20.79, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "accessed": "2026-10-25", "license": "Apache 2.0", "context": 524288, "released": "2026-10-24", "released_source": "https://huggingface.co/Nanbeige/Nanbeige4.3-33B", "note": "Dense LlamaForCausalLM, fine-tuned from Nanbeige/Nanbeige4.3-33B-Base; 33.12B is Hugging Face's safetensors total of 33,118,977,024 (BF16, about 66.26 GB). 512K context (max_position_embeddings 524288, rope_theta 1.2e8). Nanbeige and Unsloth publish no GGUF, so Q4_K_M is bartowski's (20,788,825,216 bytes); every bartowski size is in the table below. Nanbeige claims state-of-the-art agentic performance among open models around 30B (@nanbeige on X, Oct 25, 2026); that is their claim, not checked here. Release date is the Hugging Face repo creation date.", "quants": [ { "quant": "IQ2_XXS", "bytes": 10071403648, "source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "by": "bartowski" }, { "quant": "IQ2_XS", "bytes": 10634407040, "source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "by": "bartowski" }, { "quant": "IQ2_S", "bytes": 11150560384, "source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "by": "bartowski" }, { "quant": "IQ2_M", "bytes": 12187167872, "source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "by": "bartowski" }, { "quant": "Q2_K", "bytes": 13528816768, "source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "by": "bartowski" }, { "quant": "IQ3_XXS", "bytes": 14489959552, "source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "by": "bartowski" }, { "quant": "Q3_K_S", "bytes": 15211261056, "source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "by": "bartowski" }, { "quant": "IQ3_XS", "bytes": 15212665984, "source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "by": "bartowski" }, { "quant": "Q3_K_M", "bytes": 15919746176, "source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "by": "bartowski" }, { "quant": "Q3_K_L", "bytes": 17270484096, "source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "by": "bartowski" }, { "quant": "IQ3_M", "bytes": 17802005632, "source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "by": "bartowski" }, { "quant": "IQ4_XS", "bytes": 18635041920, "source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "by": "bartowski" }, { "quant": "Q4_0", "bytes": 19348057216, "source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "by": "bartowski" }, { "quant": "Q4_K_S", "bytes": 19740519552, "source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "by": "bartowski" }, { "quant": "Q4_K_M", "bytes": 20788825216, "source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "by": "bartowski" }, { "quant": "IQ4_NL", "bytes": 20871343232, "source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "by": "bartowski" }, { "quant": "Q4_1", "bytes": 21200497792, "source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "by": "bartowski" }, { "quant": "Q4_K_L", "bytes": 22344395904, "source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "by": "bartowski" }, { "quant": "Q5_K_S", "bytes": 23664311424, "source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "by": "bartowski" }, { "quant": "Q5_K_M", "bytes": 25073368192, "source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "by": "bartowski" }, { "quant": "Q6_K_S", "bytes": 27750272128, "source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "by": "bartowski" }, { "quant": "Q6_K", "bytes": 29016857728, "source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "by": "bartowski" }, { "quant": "Q6_K_L", "bytes": 30404324480, "source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "by": "bartowski" }, { "quant": "Q8_0", "bytes": 35195128960, "source": "https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF", "by": "bartowski" } ], "kv": { "layers": 52, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 524288, "config_source": "https://huggingface.co/Nanbeige/Nanbeige4.3-33B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Nanbeige/Nanbeige4.3-33B", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" } ] }, { "id": "lfm3-1.7b", "name": "LFM3-1.7B", "repo": "LiquidAI/LFM3-1.7B", "params_b": 1.73, "params_source": "https://huggingface.co/LiquidAI/LFM3-1.7B", "active_params_b": null, "active_source": null, "q4_file_gb": 1.1, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/LiquidAI/LFM3-1.7B-GGUF", "accessed": "2026-10-27", "license": "LFM Open License 1.0", "context": 262144, "released": "2026-10-27", "released_source": "https://www.liquid.ai/blog/lfm3", "note": "Liquid AI text model on the LFM3 backbone: 24 layers, 13 gated short convolutions (a small fixed-size state, no growing KV cache), 5 sliding-window attention layers (256-token window) and 6 global attention layers. 1.73B is Hugging Face's safetensors total of 1,728,801,536 (BF16). Card: 256K context; reasoning is on by default. Q4_K_M is Liquid AI's official GGUF (LFM3-1.7B-Q4_K_M.gguf, 1,099,627,808 bytes); the repo also has a quantization-aware-distilled QAD-Q4_0 build. Every size is in the table below. llama.cpp support (lfm3 architecture) was merged on Oct 27, 2026 (ggml-org/llama.cpp PR #31252) and is in tagged releases v0.9.0 (Nov 4, 2026) and v1.0.0 (Nov 13, 2026). Liquid's MLX builds wait on an mlx-lm release with LFM3 support (in progress, per the card); Transformers needs trust_remote_code. Liquid says the model can power agentic applications that used to need 9B parameters and that a 128K-token context needs 1.6 GB of KV cache; those are Liquid's claims, not checked here. LFM Open License v1.0: commercial use is licensed only while you (with your affiliates) stay under $10M annual revenue; above that, commercial use is not licensed under it. The cap does not apply to qualifying non-profits doing non-commercial or research work. Release date is Liquid's announcement.", "quants": [ { "quant": "QAD-Q4_0", "bytes": 1049615680, "source": "https://huggingface.co/LiquidAI/LFM3-1.7B-GGUF", "by": "Liquid AI" }, { "quant": "Q4_K_M", "bytes": 1099627808, "source": "https://huggingface.co/LiquidAI/LFM3-1.7B-GGUF", "by": "Liquid AI" }, { "quant": "Q5_K_M", "bytes": 1258663200, "source": "https://huggingface.co/LiquidAI/LFM3-1.7B-GGUF", "by": "Liquid AI" }, { "quant": "Q6_K", "bytes": 1427638304, "source": "https://huggingface.co/LiquidAI/LFM3-1.7B-GGUF", "by": "Liquid AI" }, { "quant": "Q8_0", "bytes": 1846238496, "source": "https://huggingface.co/LiquidAI/LFM3-1.7B-GGUF", "by": "Liquid AI" }, { "quant": "BF16", "bytes": 3466626336, "source": "https://huggingface.co/LiquidAI/LFM3-1.7B-GGUF", "by": "Liquid AI" }, { "quant": "F16", "bytes": 3466626336, "source": "https://huggingface.co/LiquidAI/LFM3-1.7B-GGUF", "by": "Liquid AI" } ], "kv": { "layers": 6, "kv_heads": 4, "head_dim": 128, "v_dim": 128, "swa_layers": 5, "swa_kv_heads": 4, "swa_head_dim": 128, "window": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/LiquidAI/LFM3-1.7B/blob/main/config.json", "config_note": "layer_types: 6 full_attention + 5 sliding_attention (256) + 13 conv layers. The conv layers keep a small fixed-size state and no KV (arch hybrid, so the speed estimate adds the recurrent-state cost); llama.cpp lfm3 puts the attention layers in its iSWA cache (src/models/lfm3.cpp, llama-memory-hybrid-iswa)" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/LiquidAI/LFM3-1.7B", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" }, { "source": "ollama", "url": "https://ollama.com/library/lfm3:1.7b", "runs": "local", "pull": "lfm3:1.7b" } ] }, { "id": "lfm3-vl-2b", "name": "LFM3-VL-2B", "repo": "LiquidAI/LFM3-VL-2B", "params_b": 1.92, "params_source": "https://huggingface.co/LiquidAI/LFM3-VL-2B", "active_params_b": null, "active_source": null, "q4_file_gb": 1.1, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/LiquidAI/LFM3-VL-2B-GGUF", "accessed": "2026-10-27", "license": "LFM Open License 1.0", "context": 128000, "released": "2026-10-27", "released_source": "https://www.liquid.ai/blog/lfm3", "note": "Liquid AI vision-language model (text, image and video in). 1.92B is Hugging Face's safetensors total of 1,918,181,888; the card splits it into a 1.73B language model, a 164M vision encoder and a 25M projector. The language model is the same 24-layer LFM3 layout as LFM3-1.7B (13 short convolutions, 5 sliding-window, 6 global attention layers) with a 128K context. Q4_K_M is Liquid AI's official GGUF (LFM3-VL-2B-Q4_K_M.gguf, 1,099,627,808 bytes, language model only); every size is in the table below. Fit ignores the vision projector (mmproj-F16 378,996,032 bytes, mmproj-Q8_0 203,031,872 bytes), which is extra if you send images. llama.cpp support (lfm3 architecture) was merged on Oct 27, 2026 (ggml-org/llama.cpp PR #31252) and is in tagged releases v0.9.0 (Nov 4, 2026) and v1.0.0 (Nov 13, 2026). The GGUF card says llama.cpp has no video input for now. Liquid's MLX builds wait on an MLX-VLM release with LFM3-VL support (in progress, per the card). LFM Open License v1.0: commercial use is licensed only while you (with your affiliates) stay under $10M annual revenue; above that, commercial use is not licensed under it. The cap does not apply to qualifying non-profits doing non-commercial or research work. Release date is Liquid's announcement; the Hugging Face repo was created Oct 19.", "quants": [ { "quant": "QAD-Q4_0", "bytes": 1049615648, "source": "https://huggingface.co/LiquidAI/LFM3-VL-2B-GGUF", "by": "Liquid AI" }, { "quant": "Q4_K_M", "bytes": 1099627808, "source": "https://huggingface.co/LiquidAI/LFM3-VL-2B-GGUF", "by": "Liquid AI" }, { "quant": "Q5_K_M", "bytes": 1258663200, "source": "https://huggingface.co/LiquidAI/LFM3-VL-2B-GGUF", "by": "Liquid AI" }, { "quant": "Q6_K", "bytes": 1427638304, "source": "https://huggingface.co/LiquidAI/LFM3-VL-2B-GGUF", "by": "Liquid AI" }, { "quant": "Q8_0", "bytes": 1846238496, "source": "https://huggingface.co/LiquidAI/LFM3-VL-2B-GGUF", "by": "Liquid AI" }, { "quant": "BF16", "bytes": 3466626336, "source": "https://huggingface.co/LiquidAI/LFM3-VL-2B-GGUF", "by": "Liquid AI" }, { "quant": "F16", "bytes": 3466626336, "source": "https://huggingface.co/LiquidAI/LFM3-VL-2B-GGUF", "by": "Liquid AI" }, { "quant": "mmproj-Q8_0", "bytes": 203031872, "source": "https://huggingface.co/LiquidAI/LFM3-VL-2B-GGUF", "by": "Liquid AI", "projector": true }, { "quant": "mmproj-F16", "bytes": 378996032, "source": "https://huggingface.co/LiquidAI/LFM3-VL-2B-GGUF", "by": "Liquid AI", "projector": true }, { "quant": "mmproj-BF16", "bytes": 378996032, "source": "https://huggingface.co/LiquidAI/LFM3-VL-2B-GGUF", "by": "Liquid AI", "projector": true } ], "kv": { "layers": 6, "kv_heads": 4, "head_dim": 128, "v_dim": 128, "swa_layers": 5, "swa_kv_heads": 4, "swa_head_dim": 128, "window": 256, "arch": "hybrid", "max_ctx": 128000, "config_source": "https://huggingface.co/LiquidAI/LFM3-VL-2B/blob/main/config.json", "config_note": "text_config layer_types: 6 full_attention + 5 sliding_attention (256) + 13 conv layers. The conv layers keep a small fixed-size state and no KV (arch hybrid, so the speed estimate adds the recurrent-state cost); llama.cpp lfm3 puts the attention layers in its iSWA cache (src/models/lfm3.cpp, llama-memory-hybrid-iswa). The vision encoder adds no KV cache" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/LiquidAI/LFM3-VL-2B", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" }, { "source": "ollama", "url": "https://ollama.com/library/lfm3-vl:2b", "runs": "local", "pull": "lfm3-vl:2b" } ] }, { "id": "mistral-large-4-1t-a52b", "name": "Mistral Large 4", "repo": "mistralai/Mistral-Large-4-1T-A52B", "params_b": 1050, "params_source": "https://docs.mistral.ai/models/mistral-large-4-0.md", "active_params_b": 52, "active_source": "https://docs.mistral.ai/models/mistral-large-4-0.md", "q4_file_gb": 605, "quant": "UD-Q4_K_XL", "q4_source": "https://unsloth.ai/docs/models/mistral-large-4", "accessed": "2026-10-29", "license": "Mistral Ecosystem Open-Weight (modified MIT)", "context": 1048576, "released": "2026-10-29", "released_source": "https://mistral.ai/news/mistral-large-4-oss/", "note": "Mistral's docs give 1.05T total and 52B active parameters. The Hugging Face model card says about 1T total and 49B active per token, or 52B counting the embedding and output layers, so both active figures are right. No plain Q4_K_M size from Unsloth, so the 4-bit size is Unsloth's UD-Q4_K_XL (604.8 GB); its smallest file is UD-IQ1_S at 233 GB. Measured runs are slower than our formula: Hugging Face measured about 15 tok/s on a 512 GB M3 Ultra (Q2_K) and 30 tok/s on four RTX Pro 6000s (IQ2_XXS, short prompts), where our formula gives 32 to 39 and about 89, so read any speed estimate for this model as roughly 2 to 3 times too high. Running it yourself is allowed for commercial use; a business offering the model as a service to third parties needs a separate license from Mistral.", "kv_skip": "Mistral Large 4's DSA sparse attention, with KV shared across layers, has no verified llama.cpp KV formula yet", "moe_skip": "no saved Hugging Face config", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/mistralai/Mistral-Large-4-1T-A52B", "uploader": "other", "note": "Account named Mistral AI; we found no Mistral page linking to it" }, { "source": "ollama", "url": "https://ollama.com/library/mistral-large-4:cloud", "runs": "cloud-only", "pull": "mistral-large-4:cloud" } ] }, { "id": "kimi-k3.1-pro", "name": "Kimi K3.1 Pro", "repo": "moonshotai/Kimi-K3.1-Pro", "params_b": 2780.11, "params_source": "https://huggingface.co/moonshotai/Kimi-K3.1-Pro", "active_params_b": 104, "active_source": "https://huggingface.co/moonshotai/Kimi-K3.1-Pro", "accessed": "2026-11-09", "license": "Kimi K3.1-Pro License", "license_source": "https://huggingface.co/moonshotai/Kimi-K3.1-Pro/blob/main/LICENSE", "context": 1048576, "released": "2026-11-02", "released_source": "https://huggingface.co/moonshotai/Kimi-K3.1-Pro", "index": true, "weights": { "bytes": 1561293160568, "shards": 96, "repo_bytes": 1561355732732, "revision": "d9e99b0792a5e93b78ea9bea62a68167e36ddd9e", "source": "https://huggingface.co/api/models/moonshotai/Kimi-K3.1-Pro?blobs=true", "note": "The whole repo, including the index and tokenizer, is 1,561,355,732,732 bytes. Commit d9e99b0792a5e93b78ea9bea62a68167e36ddd9e." }, "unknown_fit": { "lead": "Confirmed setup not yet documented.", "text": "Official weights are a 1,561.29 GB download, not the memory a machine needs. Community GGUFs now exist, including a preliminary Q4, but this site still estimates neither fit nor speed. A rough planning budget is around 2 TB of aggregate memory—approximately the capacity of four 512 GB Mac Studios. Actual requirements and support for running across those Macs remain unconfirmed." }, "note": "104B active is the model card's count of parameters used for one token. It is not the memory the checkpoint occupies, and the weight download is not that memory either. Server recipes and community uploads are not a confirmed setup for a PC or Mac on this site. Community GGUFs now exist, but their file sizes are downloads, and this site does not use them for a fit or a speed. RAM offload, an in-memory cluster, SSD streaming, and a prefill/decode split are different ways to place a model. None of them is confirmed here, and no speed or rental price is listed. A rough planning budget is around 2 TB of aggregate memory—approximately the capacity of four 512 GB Mac Studios. Actual requirements and support for running across those Macs remain unconfirmed.", "kv_skip": "Kimi K3.1 Pro's KDA and Gated DSA attention has no verified llama.cpp KV formula yet", "checks": [ { "kind": "official", "label": "Model card", "text": "Moonshot's card calls it an open-weight multimodal mixture-of-experts model: 2.8T total parameters, 104B activated per token, 93 layers (1 dense; 69 KDA and 24 Gated DSA), 896 experts with 16 selected per token plus 2 shared, a 1,048,576-token context, and text and image input through MoonViT-V2 (401M parameters). The card's quantization line is MXFP4 weights and MXFP8 activations, from quantization-aware training. Hugging Face's safetensors total is 2,780,110,363,392 parameters (2,780.11B on this page); the card rounds that to 2.8T. The 104B figure is not resident memory.", "source": "https://huggingface.co/moonshotai/Kimi-K3.1-Pro", "source_label": "Model card", "checked": "2026-11-09" }, { "kind": "official", "label": "Weight files", "text": "96 safetensors shards sum to 1,561,293,160,568 bytes (1,561.29 GB). The repo's used storage, including the index and tokenizer, is 1,561,355,732,732 bytes (1,561.36 GB). The repo was created 2026-11-02 and last modified 2026-11-05 (commit d9e99b0792a5e93b78ea9bea62a68167e36ddd9e, still the tip on 2026-11-09). Both sizes are downloads. They are not GPU memory, unified memory, or a GGUF.", "source": "https://huggingface.co/api/models/moonshotai/Kimi-K3.1-Pro?blobs=true", "source_label": "Hugging Face API", "checked": "2026-11-09" }, { "kind": "official", "label": "Which weights are MXFP4", "text": "config.json sets text_config.quantization_config (there is no top-level quantization_config) to compressed-tensors mxfp4-pack-quantized, 4-bit, group size 32. Its ignore list leaves attention, shared experts, the dense MLP, the router, lm_head, the vision tower, and the multimodal projector out of that group. kv_cache_scheme is null, so the file does not state a KV-cache format. The card's MXFP8-activations line describes the training quant, not a cache size this site can turn into a context estimate.", "source": "https://huggingface.co/moonshotai/Kimi-K3.1-Pro/blob/main/config.json", "source_label": "config.json", "checked": "2026-11-09" }, { "kind": "official", "label": "License", "text": "The weights use the Kimi K3.1-Pro License (license_name kimi-k3.1-pro). It allows use, modification, and redistribution if the copyright notice stays with the files. If you and your affiliates take more than $20 million over any 12 months from offering inference or fine-tuning as a service, you need a separate Moonshot agreement before using the weights commercially. A product with more than 100 million monthly active users, or more than $20 million in monthly revenue, must show the name Kimi K3.1-Pro. Internal use, and use through Moonshot's own products or its certified partners, is outside those two conditions. This is a summary of the license, not legal advice.", "source": "https://huggingface.co/moonshotai/Kimi-K3.1-Pro/blob/main/LICENSE", "source_label": "LICENSE", "checked": "2026-11-09" }, { "kind": "official", "label": "Hosted API", "text": "Moonshot's API quickstart names the hosted model id kimi-k3.1-pro and a 1M-token context. That is Moonshot's service, not a local setup. This page has no API price.", "source": "https://platform.kimi.ai/docs/guide/kimi-k3-1-pro-quickstart", "source_label": "API quickstart", "checked": "2026-11-09" }, { "kind": "official", "label": "Tech blog", "text": "Moonshot's tech blog introduces Kimi K3.1 Pro and a high-speed API variant, and says the full weights would be released by Nov 4, 2026. The Hugging Face repo was created on Nov 2, 2026. The blog does not name a PC, a Mac, or a measured local speed.", "source": "https://www.kimi.ai/blog/kimi-k3-1-pro", "source_label": "Moonshot tech blog", "checked": "2026-11-09" }, { "kind": "vendor", "label": "vLLM recipe", "text": "The model card links this recipe. It says the simplest NVIDIA setup it documents is tensor parallel 8, on two GB300 trays or one B300 node, and the same eight-GPU layout on one MI355X node, for the MXFP4 checkpoint. It says that was verified on NVIDIA GB300 NVL4. The page labels the variant MXFP4 1874 GB and says the routed-expert weights are about 1.56 TB on disk, with attention, shared experts, and the vision tower left in BF16. It also describes prefill/decode disaggregation and a host-memory KV tier as extra server options. Those are vLLM's statements. We have not measured resident memory or speed, and 1874 GB is not a fit in our calculator.", "source": "https://recipes.vllm.ai/moonshotai/Kimi-K3.1-Pro", "source_label": "vLLM recipe", "checked": "2026-11-09" }, { "kind": "vendor", "label": "SGLang", "text": "SGLang's November 4, 2026 day-0 post says every recipe uses a preview image and that support has not shipped in a stock SGLang release. The post says they ran their own AgentX test of the MXFP4 checkpoint on an NVIDIA GB300 NVL72, including prefill/decode splits, and it names other accelerator families. The model card links the cookbook for launch commands. The speeds in that post are SGLang's. We did not reproduce them and do not list them.", "source": "https://www.sglang.io/blog/kimi-k3.1-pro-day0-support", "source_label": "SGLang day-0 post", "checked": "2026-11-09" }, { "kind": "vendor", "label": "TokenSpeed", "text": "The model card links TokenSpeed's recipe. That page says the NVIDIA launch was validated on two GB300 nodes with 4 GPUs each, and that the context length is the checkpoint's 1,048,576 tokens. The eight-GPU MI350X or MI355X command still sets --max-model-len 65536, not the full 1,048,576. The DSpark and EAGLE draft-model flags on that page are optional speculative-decoding recipes, not a measured plain-decoding speed. That is TokenSpeed's documentation. We have not run it.", "source": "https://lightseek.org/tokenspeed/recipes/models#kimi-k3-1-pro", "source_label": "TokenSpeed recipe", "checked": "2026-11-09" }, { "kind": "vendor", "label": "Ollama cloud", "text": "Ollama's library page lists kimi-k3.1-pro:cloud. Its examples call ollama.com. That is a hosted model, not a file that runs on your machine.", "source": "https://ollama.com/library/kimi-k3.1-pro", "source_label": "Ollama library", "checked": "2026-11-09" }, { "kind": "checked", "label": "Community GGUFs", "text": "A Hugging Face search filtered to GGUF now returns several Kimi K3.1 Pro repos. Sizes below are summed from each repo's file listing and are downloads, not GPU or unified memory. DevQuasar/moonshotai.Kimi-K3.1-Pro-GGUF (card says VERY EXPERIMENTAL): MXFP4_MOE, 33 files, 1,508,897,280,480 bytes (1,508.897 GB); Q2_K, 33 files, 1,009,315,027,936 bytes (1,009.315 GB); and a BF16 mmproj of 904,448,928 bytes (0.904 GB). danielhanchen/Kimi-K3.1-Pro-GGUF-Preliminary (card title: Kimi K3.1 Pro Preliminary quants): UD-Q4_K_XL, 34 files, 1,509,020,584,672 bytes (1,509.021 GB); UD-Q8_K_XL, 34 files, 1,561,525,956,960 bytes (1,561.526 GB). Because that Q4 set is labeled Preliminary, this page does not use it as the Q4 file for the calculator. KarLexAI/Kimi-K3.1-Pro-GGUF is text only (the card says the vision encoder is not included): KLX-IQ2_XXS, 17 files, 709,876,827,040 bytes (709.877 GB); KLX-IQ1_M, 15 files, 647,869,832,960 bytes (647.870 GB); KLX-IQ1_S, 14 files, 592,677,343,328 bytes (592.677 GB). That card publishes quality and speed claims for an 8×H100 node; those are the uploader's, and we have not run them. midagedev/Kimi-K3.1-Pro-GGUF: IQ1_S, 31 files, 580,345,966,528 bytes (580.346 GB); the card says only bloomery loads it, not mainline llama.cpp, and its parameter counts do not match Moonshot's 2.8T total and 104B active. unsloth/Kimi-K3.1-Pro is still a safetensors copy of the official 96 shards (1,561,293,160,568 bytes), not a GGUF. bartowski has no Kimi K3.1 Pro repo, and mlx-community has no Kimi K3.1 Pro repo. None of these files is a confirmed setup on this site.", "source": "https://huggingface.co/api/models?search=Kimi-K3.1-Pro&filter=gguf", "source_label": "Hugging Face GGUF search", "checked": "2026-11-09" }, { "kind": "checked", "label": "llama.cpp", "text": "PR #31779, \"model: add Kimi-K3.1-Pro text model\", merged on 2026-11-06 and added text support under the architecture name kimi-k31. Vision support, PR #31980 (\"mtmd: add kimi K3.1 vision\"), merged on 2026-11-10. Both are in tagged release v1.0.0 (2026-11-13), whose release notes list Kimi-K3.1-Pro (#31779) among the new text architectures and Kimi-K3.1 vision (#31980). We have not checked whether Ollama or LM Studio can load it. A merged loader is not a confirmed setup or a speed on this site.", "source": "https://github.com/ggml-org/llama.cpp/releases/tag/v1.0.0", "source_label": "llama.cpp v1.0.0 release", "checked": "2026-11-13" }, { "kind": "checked", "label": "mlx-lm", "text": "mlx-lm main at commit 56230fa02b8ff8330daf26dc95c28480423ed5fe (2026-11-05) is still the tip of main on 2026-11-09. Its models folder has kimi_k3.py, kimi_k25.py, kimi_linear.py, and kimi_vl.py, and no kimi_k31.py. Upstream mlx-lm is not a working loader for this checkpoint.", "source": "https://github.com/ml-explore/mlx-lm/tree/56230fa02b8ff8330daf26dc95c28480423ed5fe/mlx_lm/models", "source_label": "mlx-lm models folder", "checked": "2026-11-09" }, { "kind": "community", "label": "inferencerlabs MLX upload", "text": "115 safetensors files sum to 1,561,243,578,065 bytes (1,561.244 GB); the repo was last modified 2026-11-05, after this page's earlier 112-file snapshot. The card says the build repacks the existing quantized weights and offloads the experts to storage, names a 512 GB M3 Ultra, and now publishes a speed figure. Storage offload is not the same as holding the model in GPU memory or in a Mac's unified memory. That speed figure is the uploader's, not a measurement on this site, and we have not run it.", "source": "https://huggingface.co/inferencerlabs/Kimi-K3.1-Pro-MLXse", "source_label": "inferencerlabs card", "checked": "2026-11-09" }, { "kind": "community", "label": "AinsleyMarkov 2-bit upload", "text": "95 safetensors files sum to 816,943,261,367 bytes (816.94 GB). The card says routed experts are stored at 2 bits and that the file is meant for the author's tensor-parallel MLX engine, which is not published. It also says mlx-lm has no kimi_k31 model. The 816.94 GB figure is a download size. It is not evidence that a Mac, or any machine we track, can run it.", "source": "https://huggingface.co/AinsleyMarkov/Kimi-K3.1-Pro-2bit-UVMAX", "source_label": "AinsleyMarkov card", "checked": "2026-11-09" }, { "kind": "community", "label": "pipenetwork REAP80 MLX upload", "text": "pipenetwork published REAP-pruned MLX builds that keep 179 of 896 routed experts per MoE layer and ship a custom kimi_k31.py loader; the card says the kimi_k31 architecture is not upstream in mlx-lm. The q6 build, the uploader's recommended tier, has 95 safetensors totaling 335,960,941,648 bytes (335.961 GB). The same family has a q4 build of 321,897,117,664 bytes (321.897 GB) and a q8 build of 350,024,765,392 bytes (350.025 GB). The card publishes quality and speed claims for a 512 GB M3 Ultra. Those claims are the uploader's. We have not run these files, and a pruned build with a custom loader is not a confirmed setup on this site.", "source": "https://huggingface.co/pipenetwork/Kimi-K3.1-Pro-REAP80-code-MLX-mxfp4-q6", "source_label": "pipenetwork card", "checked": "2026-11-09" }, { "kind": "community", "label": "dphn NF12 upload", "text": "96 safetensors files sum to 1,532,965,277,463 bytes (1,532.965 GB). The card describes an NF12 compression of the checkpoint's BF16 tensors, aimed at an 8-GPU B300-class server with the author's runtime. It is not a GGUF or an mlx-lm build, and we have not run it.", "source": "https://huggingface.co/dphn/Kimi-K3.1-Pro-NF12", "source_label": "dphn card", "checked": "2026-11-09" }, { "kind": "unknown", "label": "Not documented", "text": "Unknown, and not estimated here: how much memory a runtime keeps resident, KV-cache bytes per token, any measured speed, a tested PC or Mac in our catalog, RAM offload of the weights, an in-memory cluster of machines we list, and a current rental price. Prefill/decode splits appear only as options inside the server recipes above. They are not a confirmed setup for one machine.", "source": "https://huggingface.co/moonshotai/Kimi-K3.1-Pro", "source_label": "Model card does not give this", "checked": "2026-11-09" } ], "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/moonshotai/Kimi-K3.1-Pro", "uploader": "official" }, { "source": "ollama", "url": "https://ollama.com/library/kimi-k3.1-pro:cloud", "runs": "cloud-only", "pull": "kimi-k3.1-pro:cloud" } ] }, { "id": "longcat-2.5", "name": "LongCat-2.5", "repo": "meituan-longcat/LongCat-2.5", "params_b": 1779.75, "params_source": "https://huggingface.co/meituan-longcat/LongCat-2.5", "active_params_b": 48, "active_source": "https://huggingface.co/meituan-longcat/LongCat-2.5", "accessed": "2026-11-09", "license": "MIT", "license_source": "https://huggingface.co/meituan-longcat/LongCat-2.5/blob/main/LICENSE", "context": 524288, "released": "2026-11-04", "released_source": "https://huggingface.co/meituan-longcat/LongCat-2.5", "index": true, "weights": { "bytes": 3561974336256, "shards": 210, "repo_bytes": 3561989945589, "revision": "6fc4844cc9455bc45b84871cbc9137b3286f0fc2", "source": "https://huggingface.co/api/models/meituan-longcat/LongCat-2.5?blobs=true", "note": "The 210 files are 208 model shards (3,553,858,076,424 bytes) and two multimodal model_mm files (8,116,259,832 bytes). The repo's used storage is 3,561,989,945,589 bytes. Commit 6fc4844cc9455bc45b84871cbc9137b3286f0fc2." }, "unknown_fit": { "lead": "Confirmed setup not yet documented.", "text": "MIT open weights; 3,561.97 GB BF16 download, not memory. Cluster-scale NPU recipe only, no PC/Mac. No Q4, no fit or speed." }, "note": "48B active is the model card's count of parameters used for one token. It is not the memory the checkpoint occupies, and the weight download is not that memory either. The card claims 1.6 trillion total parameters, and Hugging Face's safetensors metadata sums to 1,779,752,648,320 (1,779.75B on this page). Those are two published figures. This page lists both and does not guess at the difference. config.json sets a 524,288-token context, which is the window shown here. Some evaluation footnotes and the NPU launch script use about 1 million tokens, and this page does not treat that as a confirmed window. The official W4A4 checkpoint is 1,336.01 GB of safetensors for NPU serving. It is not a GGUF and not a file the calculator can score. The 194-NPU Atlas A2 layout in Meituan's serving guide is one documented recipe, not a minimum. We have not run LongCat-2.5, and no speed or rental price is listed.", "kv_skip": "LongCat-2.5's MLA and related cache layout has no verified llama.cpp KV formula yet", "related": [ { "href": "/models/", "k": "Catalog", "t": "Every model in the catalog" }, { "href": "/methodology/#speed", "k": "Method", "t": "Why this site estimates no fit or speed without a Q4 file" }, { "href": "/models/kimi-k3.1-pro/", "k": "Model", "t": "Kimi K3.1 Pro, another large open MoE with no Q4 file in the calculator" } ], "checks": [ { "kind": "official", "label": "Model card", "text": "Meituan's card describes LongCat-2.5 as a native multimodal mixture-of-experts model and claims 1.6 trillion total parameters with about 48 billion activated per token. It lists the license as MIT. Its deployment section documents NPU serving through SGLang-FluentLLM and nothing else: no NVIDIA GPU, llama.cpp, MLX, or vLLM recipe. Hugging Face's safetensors metadata totals 1,779,752,648,320 parameters (1,779.75B on this page; the Hugging Face page rounds it to 1.8T). The card's 1.6T and the metadata's 1.78T are two published figures, and this page does not guess why they differ. The 48B figure is not resident memory.", "source": "https://huggingface.co/meituan-longcat/LongCat-2.5", "source_label": "Model card", "checked": "2026-11-09" }, { "kind": "official", "label": "Weight files", "text": "210 safetensors files sum to 3,561,974,336,256 bytes (3,561.97 GB): 208 complete model shards (model_*-of-00208) at 3,553,858,076,424 bytes and two model_mm files at 8,116,259,832 bytes. The repo's used storage is 3,561,989,945,589 bytes, and the full file tree is 3,562,006,190,340 bytes. The metadata splits the parameters into 1,778,523,785,600 BF16 and 1,228,862,720 F32. The repo was created 2026-11-04 and last modified 2026-11-09 (commit 6fc4844cc9455bc45b84871cbc9137b3286f0fc2), the day of the public announcement. We read the sizes from the Hugging Face API and did not download the files. They are download sizes. They are not GPU memory, unified memory, or a GGUF.", "source": "https://huggingface.co/api/models/meituan-longcat/LongCat-2.5?blobs=true", "source_label": "Hugging Face API", "checked": "2026-11-09" }, { "kind": "official", "label": "Architecture config", "text": "config.json sets model_type pro_v and architecture LongcatCausalLM: 38 layers with MLA attention, 768 routed experts with 12 selected per token (moe_topk), and 128 identity (zero-cost) experts. max_position_embeddings is 524,288, the context shown on this page. Some evaluation footnotes and the NPU launch script use about 1 million tokens. This page does not treat that as the model's confirmed window.", "source": "https://huggingface.co/meituan-longcat/LongCat-2.5/blob/main/config.json", "source_label": "config.json", "checked": "2026-11-09" }, { "kind": "official", "label": "License", "text": "The LICENSE file is the standard MIT license, Copyright (c) 2026 Meituan. It allows use, copying, modification, distribution, and sale, as long as the copyright and permission notice stay with the copies. This is a summary of the license, not legal advice.", "source": "https://huggingface.co/meituan-longcat/LongCat-2.5/blob/main/LICENSE", "source_label": "LICENSE", "checked": "2026-11-09" }, { "kind": "official", "label": "W4A4 weight files", "text": "Meituan also publishes LongCat-2.5-W4A4. Its 210 safetensors files sum to 1,336,010,166,464 bytes (1,336.01 GB), with all 208 main shards present. The full file tree is 1,336,050,649,636 bytes and used storage is 1,336,042,442,734 bytes; the repo was last modified 2026-11-09. This is the checkpoint the NPU guide serves. It is not a GGUF and not a verified fit on any consumer machine, and this site does not use its size as a Q4 file.", "source": "https://huggingface.co/meituan-longcat/LongCat-2.5-W4A4", "source_label": "W4A4 repo", "checked": "2026-11-09" }, { "kind": "official", "label": "Hosted chat", "text": "Meituan's LongCat site offers a chat interface. That is Meituan's service, not a local setup, and this page lists no price for it.", "source": "https://longcat.ai/", "source_label": "longcat.ai", "checked": "2026-11-09" }, { "kind": "vendor", "label": "SGLang-FluentLLM NPU guide", "text": "The model card points to SGLang-FluentLLM. Its longcat-2.5-npu branch has the documented LongCat-2.5 recipe: the W4A4 checkpoint on Atlas A2 with an Encoder-Prefill-Decode split across 194 NPUs (2 encoder, 64 prefill, 128 decode). That is one documented cluster recipe. It is not a minimum hardware requirement and not a memory floor. The project's main README mentions NVIDIA H800 and H20 testing, but its Quick Start covers LongCat-Flash-Lite, and no LongCat-2.5 GPU recipe is documented. We have not run either.", "source": "https://github.com/meituan-longcat/SGLang-FluentLLM/blob/longcat-2.5-npu/npu_test/ReadMe.md", "source_label": "NPU guide", "checked": "2026-11-09" }, { "kind": "checked", "label": "Loader code and encoder weights", "text": "config.json's auto_map names configuration_longcat_oe.py and modeling_longcat_oe.py, and the repo is tagged custom_code, but neither the Hugging Face nor the ModelScope file tree has any .py file. The NPU guide loads weights from a LongCat-2.5-Encoder path, and we found no public Hugging Face or ModelScope repo under that name. Meituan has no LongCat-2.5 model repo on GitHub; the serving code is in SGLang-FluentLLM.", "source": "https://huggingface.co/meituan-longcat/LongCat-2.5/tree/main", "source_label": "Hugging Face file tree", "checked": "2026-11-09" }, { "kind": "checked", "label": "No GGUF, and no mlx-community port", "text": "Hugging Face searches for LongCat-2.5 filtered to GGUF and to MLX returned no models, and mlx-community has no LongCat-2.5 repo. Ports of the older LongCat-Flash models are different models and are not listed here.", "source": "https://huggingface.co/api/models?search=LongCat-2.5&filter=gguf", "source_label": "Hugging Face GGUF search", "checked": "2026-11-09" }, { "kind": "checked", "label": "llama.cpp", "text": "llama.cpp master has no longcat or pro_v architecture. Open pull requests add LongCat-Flash and LongCat-Flash-Lite, which are different models, and none of them is merged.", "source": "https://github.com/ggml-org/llama.cpp/blob/master/src/llama-arch.cpp", "source_label": "llama.cpp architectures", "checked": "2026-11-09" }, { "kind": "checked", "label": "mlx-lm", "text": "mlx-lm's model folder has longcat_flash.py and longcat_flash_ngram.py, and no file for LongCat-2.5's pro_v architecture.", "source": "https://github.com/ml-explore/mlx-lm/tree/main/mlx_lm/models", "source_label": "mlx-lm models", "checked": "2026-11-09" }, { "kind": "unknown", "label": "Not documented", "text": "Unknown, and not estimated here: how much memory a runtime keeps resident, KV-cache bytes per token, any measured tokens per second, a LongCat-2.5 GPU recipe, a tested PC or Mac in our catalog, and a rental price.", "source": "https://huggingface.co/meituan-longcat/LongCat-2.5", "source_label": "Model card does not give this", "checked": "2026-11-09" } ], "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/meituan-longcat/LongCat-2.5", "uploader": "official" } ] }, { "id": "kvark-sparkle-1.1-27b", "name": "Kvark Sparkle 1.1 27B", "repo": "Egzakta/Kvark-Sparkle-1.1-27B", "params_b": 27.78, "params_source": "https://huggingface.co/Egzakta/Kvark-Sparkle-1.1-27B", "active_params_b": null, "active_source": null, "q4_file_gb": 16.81, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/Egzakta/Kvark-Sparkle-1.1-27B-GGUF", "accessed": "2026-11-11", "license": "Apache 2.0", "license_source": "https://huggingface.co/Egzakta/Kvark-Sparkle-1.1-27B/blob/main/LICENSE", "context": 262144, "released": "2026-11-10", "released_source": "https://huggingface.co/Egzakta/Kvark-Sparkle-1.1-27B", "note": "Egzakta's fine-tune of Kvark Sparkle 1.0 (itself built on Qwen3.8-27B) for document work: searching a document base, answering from what it finds, and declining when the answer isn't there. It was trained with a \"stop rule\" appended to the system prompt, and the card says it works in English and Serbian. Thinking is on by default. The card says the architecture is Qwen3.8-27B, unchanged; vision uses the base model's encoder and was not separately evaluated. Q4_K_M is Egzakta's official GGUF (Kvark-Sparkle-1.1-27B-Q4_K_M.gguf, 16,810,715,744 bytes); every size is in the table below. Fit leaves out the 0.93 GB vision projector, which is extra if you send images. The checkpoint and every GGUF keep a 1-layer MTP head, so speculative decoding (MTP drafting) is possible; that is not a speed figure, and the speed estimates on this page do not assume it. Release date is the Hugging Face repo creation date.", "quants": [ { "quant": "IQ4_XS", "bytes": 15309040224, "source": "https://huggingface.co/Egzakta/Kvark-Sparkle-1.1-27B-GGUF", "by": "Egzakta", "file": "Kvark-Sparkle-1.1-27B-IQ4_XS.gguf" }, { "quant": "Q4_K_M", "bytes": 16810715744, "source": "https://huggingface.co/Egzakta/Kvark-Sparkle-1.1-27B-GGUF", "by": "Egzakta", "file": "Kvark-Sparkle-1.1-27B-Q4_K_M.gguf" }, { "quant": "Q5_K_M", "bytes": 19535702624, "source": "https://huggingface.co/Egzakta/Kvark-Sparkle-1.1-27B-GGUF", "by": "Egzakta", "file": "Kvark-Sparkle-1.1-27B-Q5_K_M.gguf" }, { "quant": "Q6_K", "bytes": 22431001184, "source": "https://huggingface.co/Egzakta/Kvark-Sparkle-1.1-27B-GGUF", "by": "Egzakta", "file": "Kvark-Sparkle-1.1-27B-Q6_K.gguf" }, { "quant": "Q8_0", "bytes": 29047085664, "source": "https://huggingface.co/Egzakta/Kvark-Sparkle-1.1-27B-GGUF", "by": "Egzakta", "file": "Kvark-Sparkle-1.1-27B-Q8_0.gguf" }, { "quant": "BF16", "bytes": 54657735008, "source": "https://huggingface.co/Egzakta/Kvark-Sparkle-1.1-27B-GGUF", "by": "Egzakta", "file": "Kvark-Sparkle-1.1-27B-BF16.gguf" }, { "quant": "mmproj-F16", "bytes": 927606816, "source": "https://huggingface.co/Egzakta/Kvark-Sparkle-1.1-27B-GGUF", "by": "Egzakta", "file": "mmproj-Kvark-Sparkle-1.1-27B-F16.gguf", "projector": true } ], "weights": { "bytes": 55563007192, "shards": 13, "revision": "cb02a553030910573c450940eb12477d27370720", "source": "https://huggingface.co/api/models/Egzakta/Kvark-Sparkle-1.1-27B?blobs=true", "note": "BF16 safetensors: 12 model shards plus model-mtp.safetensors (849,400,424 bytes). Commit cb02a553030910573c450940eb12477d27370720." }, "checks": [ { "kind": "official", "label": "Model card", "text": "Egzakta's card says \"Architecture: Qwen3.8-27B, unchanged\"; the base model is Egzakta/Kvark-Sparkle-1.0-27B, a fine-tune of Qwen3.8-27B. Context: \"262,144 tokens native; evaluated up to 64k\". config.json matches Qwen3.8-27B's: 64 layers (16 full attention), 4 KV heads, head size 256, the same context, 1 MTP layer and a vision tower. Hugging Face's safetensors total is 27,781,427,952 parameters (BF16). The card gives the KV cache as \"64 KiB per token in BF16, 32 KiB in FP8. Only the 16 full-attention layers store it.\" That agrees with the 64 KB per token this page derives from config.json.", "source": "https://huggingface.co/Egzakta/Kvark-Sparkle-1.1-27B", "source_label": "Model card", "checked": "2026-11-11" }, { "kind": "official", "label": "License", "text": "Apache 2.0. The card: \"Apache 2.0, inherited from Qwen3.8-27B... Commercial use is permitted, with no revenue cap.\" This is a summary of the license, not legal advice.", "source": "https://huggingface.co/Egzakta/Kvark-Sparkle-1.1-27B/blob/main/LICENSE", "source_label": "LICENSE", "checked": "2026-11-11" }, { "kind": "official", "label": "Weight files", "text": "13 BF16 safetensors files sum to 55,563,007,192 bytes (55.56 GB): 12 model shards and model-mtp.safetensors (849,400,424 bytes), at commit cb02a553030910573c450940eb12477d27370720. Egzakta's FP8 checkpoint for vLLM (Egzakta/Kvark-Sparkle-1.1-27B-FP8) is 30,866,867,288 bytes. These are downloads, not the memory a runtime needs.", "source": "https://huggingface.co/api/models/Egzakta/Kvark-Sparkle-1.1-27B?blobs=true", "source_label": "Hugging Face API", "checked": "2026-11-11" }, { "kind": "official", "label": "GGUF files", "text": "Egzakta's own GGUF repo (commit 5cca6274422347e4dc67fd0e4cc4049aeff79e40) uses the qwen35 architecture, so stock llama.cpp loads it. The card says every file keeps the MTP head and calls Q4_K_M \"recommended for 24 GB cards\" (Egzakta's words). There is no bartowski repo. Every size is in the table below.", "source": "https://huggingface.co/Egzakta/Kvark-Sparkle-1.1-27B-GGUF", "source_label": "GGUF repo", "checked": "2026-11-11" }, { "kind": "vendor", "label": "Memory table (Egzakta's figures)", "text": "Egzakta's GGUF card lists the llama.cpp KV cache as 2 GiB at f16 or 1.1 GiB at q8_0 for 32,768 tokens; 4 or 2.1 GiB at 65,536; 8 or 4.3 GiB at 131,072; 16 or 8.5 GiB at 262,144; plus a linear-attention state under 0.2 GB at any context. Those are Egzakta's figures. The context estimates on this page come from our own formula.", "source": "https://huggingface.co/Egzakta/Kvark-Sparkle-1.1-27B-GGUF", "source_label": "GGUF card", "checked": "2026-11-11" }, { "kind": "vendor", "label": "MTP drafting and speed test", "text": "Egzakta's llama.cpp command uses --spec-type draft-mtp --spec-draft-n-max 3, which drafts tokens with the MTP head. That makes speculative decoding possible; it is not a speed figure. Egzakta's own test, with MTP drafting on; time per answer, not tok/s: one user, 1× H100, vLLM 0.32.0, FP8 weights, 1 MTP draft token, 20 fixed questions, MTP drafting on for both models. Median time per answer was 32.9 s for Sparkle 1.1 and 40.8 s for Qwen3.8-27B. We have not run it, and it is not a speed estimate on this page.", "source": "https://huggingface.co/Egzakta/Kvark-Sparkle-1.1-27B", "source_label": "Model card", "checked": "2026-11-11" }, { "kind": "vendor", "label": "Benchmark scores", "text": "Egzakta's card reports public-benchmark scores against Qwen3.8-27B and reasoning-token counts. Those are Egzakta's own reported scores, not checked by us, and this page does not list them.", "source": "https://huggingface.co/Egzakta/Kvark-Sparkle-1.1-27B", "source_label": "Model card", "checked": "2026-11-11" }, { "kind": "checked", "label": "stop_system.txt", "text": "The card tells users to download stop_system.txt from the repo, but at commit cb02a553 that file is not in the repo (404). The rule's full text is quoted on the card.", "source": "https://huggingface.co/Egzakta/Kvark-Sparkle-1.1-27B/tree/cb02a553030910573c450940eb12477d27370720", "source_label": "Hugging Face file tree", "checked": "2026-11-11" }, { "kind": "community", "label": "Community GGUFs", "text": "mradermacher publishes static and imatrix GGUFs (mradermacher/Kvark-Sparkle-1.1-27B-GGUF and mradermacher/Kvark-Sparkle-1.1-27B-i1-GGUF). We did not size them; this page uses Egzakta's own files.", "source": "https://huggingface.co/mradermacher/Kvark-Sparkle-1.1-27B-GGUF", "source_label": "mradermacher repo", "checked": "2026-11-11" }, { "kind": "checked", "label": "MLX", "text": "No MLX build of Kvark Sparkle 1.1 on Hugging Face as of 2026-11-11. Community MLX builds exist only for Sparkle 1.0, a different checkpoint.", "source": "https://huggingface.co/models?search=Kvark-Sparkle-1.1&library=mlx", "source_label": "Hugging Face MLX search", "checked": "2026-11-11" } ], "related": [ { "href": "/models/qwen3.8-27b/", "k": "Model", "t": "Base model family: Qwen3.8-27B" } ], "kv": { "layers": 16, "kv_heads": 4, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/Egzakta/Kvark-Sparkle-1.1-27B/blob/main/config.json" } }, { "id": "zoomgain-4b", "name": "ZoomGain-4B", "repo": "inclusionAI/ZoomGain-4B", "params_b": 4.54, "params_source": "https://huggingface.co/inclusionAI/ZoomGain-4B", "active_params_b": null, "active_source": null, "q4_file_gb": 2.71, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/inclusionAI/ZoomGain-4B-GGUF", "accessed": "2026-11-11", "license": "Not stated by inclusionAI", "license_source": "https://huggingface.co/inclusionAI/ZoomGain-4B", "context": 262144, "released": "2026-11-10", "released_source": "https://huggingface.co/inclusionAI/ZoomGain-4B", "note": "inclusionAI vision-language model, RL-trained from Qwen3.5-4B on inclusionAI's ZoomGain data for fine-grained visual perception: details that only show when you zoom in, such as attributes, counting and OCR (paper: arXiv 2611.14076). Its config matches Qwen3.5-4B's except that the MTP layer is removed (mtp_num_hidden_layers 0), so there is no built-in MTP drafting. Image input needs the mmproj file, which the fit leaves out. Q4_K_M is inclusionAI's official GGUF (ZoomGain-4B-Q4_K_M.gguf, 2,708,804,160 bytes); every size is in the table below. License: the Hugging Face card and repo state none for these weights, so confirm the terms with inclusionAI before commercial use (details below). Release date is the Hugging Face repo creation date.", "quants": [ { "quant": "Q4_K_M", "bytes": 2708804160, "source": "https://huggingface.co/inclusionAI/ZoomGain-4B-GGUF", "by": "inclusionAI", "file": "ZoomGain-4B-Q4_K_M.gguf" }, { "quant": "Q8_0", "bytes": 4482402880, "source": "https://huggingface.co/inclusionAI/ZoomGain-4B-GGUF", "by": "inclusionAI", "file": "ZoomGain-4B-Q8_0.gguf" }, { "quant": "F16", "bytes": 8424393280, "source": "https://huggingface.co/inclusionAI/ZoomGain-4B-GGUF", "by": "inclusionAI", "file": "ZoomGain-4B-F16.gguf" }, { "quant": "Q4_K_M", "bytes": 2729642816, "source": "https://huggingface.co/bartowski/ZoomGain-4B-GGUF", "by": "bartowski", "file": "ZoomGain-4B-Q4_K_M.gguf" }, { "quant": "Q6_K", "bytes": 3689878336, "source": "https://huggingface.co/bartowski/ZoomGain-4B-GGUF", "by": "bartowski", "file": "ZoomGain-4B-Q6_K.gguf" }, { "quant": "Q8_0", "bytes": 4493953856, "source": "https://huggingface.co/bartowski/ZoomGain-4B-GGUF", "by": "bartowski", "file": "ZoomGain-4B-Q8_0.gguf" }, { "quant": "mmproj-F16", "bytes": 672423040, "source": "https://huggingface.co/inclusionAI/ZoomGain-4B-GGUF", "by": "inclusionAI", "file": "mmproj-ZoomGain-4B-F16.gguf", "projector": true }, { "quant": "mmproj-Q8_0", "bytes": 363748480, "source": "https://huggingface.co/inclusionAI/ZoomGain-4B-GGUF", "by": "inclusionAI", "file": "mmproj-ZoomGain-4B-Q8_0.gguf", "projector": true }, { "quant": "mmproj-Q4_K_M", "bytes": 234806400, "source": "https://huggingface.co/inclusionAI/ZoomGain-4B-GGUF", "by": "inclusionAI", "file": "mmproj-ZoomGain-4B-Q4_K_M.gguf", "projector": true } ], "weights": { "bytes": 9078620536, "shards": 1, "revision": "2ac9e28f1de993b4d7b6ebeda1368ff2e6982a71", "source": "https://huggingface.co/api/models/inclusionAI/ZoomGain-4B?blobs=true", "note": "One BF16 safetensors file. Commit 2ac9e28f1de993b4d7b6ebeda1368ff2e6982a71." }, "checks": [ { "kind": "unknown", "label": "License not stated", "text": "The Hugging Face model card and repo state no license for the ZoomGain-4B weights: there is no license metadata and no LICENSE file. The GitHub repo's code is Apache 2.0, and its README says \"See the Hugging Face cards for the terms of each dataset and checkpoint\", which state none. The base model, Qwen3.5-4B, is Apache 2.0, but that does not set the terms for this checkpoint. Confirm the terms with inclusionAI before commercial use.", "source": "https://huggingface.co/inclusionAI/ZoomGain-4B", "source_label": "Model card", "checked": "2026-11-11" }, { "kind": "official", "label": "Code and paper", "text": "inclusionAI's code is at github.com/inclusionAI/ZoomGain (the code is Apache 2.0) and the paper is arXiv 2611.14076. The README points to the Hugging Face cards for checkpoint terms.", "source": "https://github.com/inclusionAI/ZoomGain", "source_label": "GitHub repo", "checked": "2026-11-11" }, { "kind": "official", "label": "Weight files", "text": "One BF16 safetensors file of 9,078,620,536 bytes (9.08 GB) at commit 2ac9e28f1de993b4d7b6ebeda1368ff2e6982a71. Hugging Face's safetensors total is 4,539,265,536 parameters. This is a download, not the memory a runtime needs.", "source": "https://huggingface.co/api/models/inclusionAI/ZoomGain-4B?blobs=true", "source_label": "Hugging Face API", "checked": "2026-11-11" }, { "kind": "official", "label": "GGUF files", "text": "inclusionAI's own GGUF repo (commit 851a0074adb9b89f2c043a1c55dbf155b9e0acb2) uses the qwen35 architecture. It has Q4_K_M, Q8_0 and F16 language-model files and three mmproj projectors (F16, Q8_0, Q4_K_M) for image input. Every size is in the table below.", "source": "https://huggingface.co/inclusionAI/ZoomGain-4B-GGUF", "source_label": "GGUF repo", "checked": "2026-11-11" }, { "kind": "community", "label": "bartowski GGUFs", "text": "bartowski/ZoomGain-4B-GGUF (commit aa3e7c80) has Q4_K_M, Q6_K and Q8_0 files; their sizes are in the table below. The fit on this page uses inclusionAI's own Q4_K_M.", "source": "https://huggingface.co/bartowski/ZoomGain-4B-GGUF", "source_label": "bartowski repo", "checked": "2026-11-11" }, { "kind": "checked", "label": "MLX", "text": "No MLX build of ZoomGain-4B on Hugging Face as of 2026-11-11.", "source": "https://huggingface.co/models?search=ZoomGain-4B&library=mlx", "source_label": "Hugging Face MLX search", "checked": "2026-11-11" } ], "related": [ { "href": "/models/qwen3.5-4b/", "k": "Model", "t": "Base model: Qwen3.5-4B" }, { "href": "/models/zoomgain-9b/", "k": "Model", "t": "ZoomGain-9B, the larger size" } ], "kv": { "layers": 8, "kv_heads": 4, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/inclusionAI/ZoomGain-4B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/inclusionAI/ZoomGain-4B", "uploader": "official" } ] }, { "id": "zoomgain-9b", "name": "ZoomGain-9B", "repo": "inclusionAI/ZoomGain-9B", "params_b": 9.41, "params_source": "https://huggingface.co/inclusionAI/ZoomGain-9B", "active_params_b": null, "active_source": null, "q4_file_gb": 5.63, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/inclusionAI/ZoomGain-9B-GGUF", "accessed": "2026-11-11", "license": "Not stated by inclusionAI", "license_source": "https://huggingface.co/inclusionAI/ZoomGain-9B", "context": 262144, "released": "2026-11-10", "released_source": "https://huggingface.co/inclusionAI/ZoomGain-9B", "note": "inclusionAI vision-language model, RL-trained from Qwen3.5-9B on inclusionAI's ZoomGain data for fine-grained visual perception: details that only show when you zoom in, such as attributes, counting and OCR (paper: arXiv 2611.14076). Its config matches Qwen3.5-9B's except that the MTP layer is removed (mtp_num_hidden_layers 0), so there is no built-in MTP drafting. Image input needs the mmproj file, which the fit leaves out. Q4_K_M is inclusionAI's official GGUF (ZoomGain-9B-Q4_K_M.gguf, 5,629,108,832 bytes); every size is in the table below. License: the Hugging Face card and repo state none for these weights, so confirm the terms with inclusionAI before commercial use (details below). Release date is the Hugging Face repo creation date.", "quants": [ { "quant": "Q4_K_M", "bytes": 5629108832, "source": "https://huggingface.co/inclusionAI/ZoomGain-9B-GGUF", "by": "inclusionAI", "file": "ZoomGain-9B-Q4_K_M.gguf" }, { "quant": "Q8_0", "bytes": 9527501408, "source": "https://huggingface.co/inclusionAI/ZoomGain-9B-GGUF", "by": "inclusionAI", "file": "ZoomGain-9B-Q8_0.gguf" }, { "quant": "F16", "bytes": 17920696928, "source": "https://huggingface.co/inclusionAI/ZoomGain-9B-GGUF", "by": "inclusionAI", "file": "ZoomGain-9B-F16.gguf" }, { "quant": "mmproj-F16", "bytes": 918165536, "source": "https://huggingface.co/inclusionAI/ZoomGain-9B-GGUF", "by": "inclusionAI", "file": "mmproj-ZoomGain-9B-F16.gguf", "projector": true }, { "quant": "mmproj-Q8_0", "bytes": 620690336, "source": "https://huggingface.co/inclusionAI/ZoomGain-9B-GGUF", "by": "inclusionAI", "file": "mmproj-ZoomGain-9B-Q8_0.gguf", "projector": true }, { "quant": "mmproj-Q4_K_M", "bytes": 516095648, "source": "https://huggingface.co/inclusionAI/ZoomGain-9B-GGUF", "by": "inclusionAI", "file": "mmproj-ZoomGain-9B-Q4_K_M.gguf", "projector": true } ], "weights": { "bytes": 18819722392, "shards": 1, "revision": "b4550ef2e5a30b33c5133278036a682197489145", "source": "https://huggingface.co/api/models/inclusionAI/ZoomGain-9B?blobs=true", "note": "One BF16 safetensors file. Commit b4550ef2e5a30b33c5133278036a682197489145." }, "checks": [ { "kind": "unknown", "label": "License not stated", "text": "The Hugging Face model card and repo state no license for the ZoomGain-9B weights: there is no license metadata and no LICENSE file. The GitHub repo's code is Apache 2.0, and its README says \"See the Hugging Face cards for the terms of each dataset and checkpoint\", which state none. The base model, Qwen3.5-9B, is Apache 2.0, but that does not set the terms for this checkpoint. Confirm the terms with inclusionAI before commercial use.", "source": "https://huggingface.co/inclusionAI/ZoomGain-9B", "source_label": "Model card", "checked": "2026-11-11" }, { "kind": "official", "label": "Code and paper", "text": "inclusionAI's code is at github.com/inclusionAI/ZoomGain (the code is Apache 2.0) and the paper is arXiv 2611.14076. The README points to the Hugging Face cards for checkpoint terms.", "source": "https://github.com/inclusionAI/ZoomGain", "source_label": "GitHub repo", "checked": "2026-11-11" }, { "kind": "official", "label": "Weight files", "text": "One BF16 safetensors file of 18,819,722,392 bytes (18.82 GB) at commit b4550ef2e5a30b33c5133278036a682197489145. Hugging Face's safetensors total is 9,409,813,744 parameters. This is a download, not the memory a runtime needs.", "source": "https://huggingface.co/api/models/inclusionAI/ZoomGain-9B?blobs=true", "source_label": "Hugging Face API", "checked": "2026-11-11" }, { "kind": "official", "label": "GGUF files", "text": "inclusionAI's own GGUF repo (commit 3c25485839f778c068f8813fd6284a994b8ff6f6) uses the qwen35 architecture. It has Q4_K_M, Q8_0 and F16 language-model files and three mmproj projectors (F16, Q8_0, Q4_K_M) for image input. Every size is in the table below.", "source": "https://huggingface.co/inclusionAI/ZoomGain-9B-GGUF", "source_label": "GGUF repo", "checked": "2026-11-11" }, { "kind": "vendor", "label": "inclusionAI's claim", "text": "inclusionAI's card says ZoomGain-9B \"reached superior performance on multiple fine-grained perception and general perception benchmarks\". That is inclusionAI's claim, not checked by us. The scores are only in a table image, so this page lists none.", "source": "https://huggingface.co/inclusionAI/ZoomGain-9B", "source_label": "Model card", "checked": "2026-11-11" }, { "kind": "checked", "label": "Community GGUFs", "text": "No community GGUF of ZoomGain-9B on Hugging Face as of 2026-11-11; this page uses inclusionAI's own files.", "source": "https://huggingface.co/models?search=ZoomGain-9B&library=gguf", "source_label": "Hugging Face GGUF search", "checked": "2026-11-11" }, { "kind": "checked", "label": "MLX", "text": "No MLX build of ZoomGain-9B on Hugging Face as of 2026-11-11.", "source": "https://huggingface.co/models?search=ZoomGain-9B&library=mlx", "source_label": "Hugging Face MLX search", "checked": "2026-11-11" } ], "related": [ { "href": "/models/qwen3.5-9b/", "k": "Model", "t": "Base model: Qwen3.5-9B" }, { "href": "/models/zoomgain-4b/", "k": "Model", "t": "ZoomGain-4B, the smaller size" } ], "kv": { "layers": 8, "kv_heads": 4, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/inclusionAI/ZoomGain-9B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/inclusionAI/ZoomGain-9B", "uploader": "official" } ] }, { "id": "botpsy-0.8b", "name": "BotPsy-0.8B", "repo": "qvac/BotPsy-0.8B", "params_b": 0.85, "params_source": "https://huggingface.co/qvac/BotPsy-0.8B", "active_params_b": null, "active_source": null, "q4_file_gb": 0.53, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/qvac/BotPsy-0.8B-GGUFs", "accessed": "2026-11-11", "license": "Apache 2.0", "license_source": "https://huggingface.co/qvac/BotPsy-0.8B/blob/main/LICENSE", "context": 262144, "released": "2026-11-10", "released_source": "https://huggingface.co/qvac/BotPsy-0.8B", "note": "A 0.8B tool-calling (function-calling) assistant for on-device agents, from Tether AI Research (Tether Data, S.A. de C.V.), published under the qvac org. Fine-tuned from Qwen3.5-0.8B (huggingface.co/Qwen/Qwen3.5-0.8B), which has no page here. Tether says it suits tasks of low to medium difficulty, that it thinks before answering (enable_thinking=True; the chat template defaults to non-thinking), that English is recommended, and that it \"is not intended for safety-critical or high-stakes automated decisions, or for applications involving money, personal information, health and other sensitive topics.\" It runs on Transformers, vLLM, llama.cpp and Tether's QVAC SDK. Text only, with no projector. The base model's MTP layers are not included (mtp_num_hidden_layers 0), so there is no MTP drafting. Q4_K_M is Tether's official imatrix-calibrated GGUF (botpsy-0.8b-q4_k_m-imat.gguf, 529,296,960 bytes); every size is in the table below. Release date is the Hugging Face repo creation date.", "quants": [ { "quant": "Q4_K_M", "bytes": 529296960, "source": "https://huggingface.co/qvac/BotPsy-0.8B-GGUFs", "by": "Tether", "file": "botpsy-0.8b-q4_k_m-imat.gguf" }, { "quant": "Q8_0", "bytes": 811843040, "source": "https://huggingface.co/qvac/BotPsy-0.8B-GGUFs", "by": "Tether", "file": "botpsy-0.8b-q8_0.gguf" }, { "quant": "BF16", "bytes": 1516744160, "source": "https://huggingface.co/qvac/BotPsy-0.8B-GGUFs", "by": "Tether", "file": "botpsy-0.8b-bf16.gguf" } ], "weights": { "bytes": 1706030528, "shards": 1, "revision": "24561e9681643eaf807fc28b0875210cb17f89c1", "source": "https://huggingface.co/api/models/qvac/BotPsy-0.8B?blobs=true", "note": "One BF16 safetensors file. Commit 24561e9681643eaf807fc28b0875210cb17f89c1." }, "checks": [ { "kind": "official", "label": "License", "text": "Apache 2.0, from the LICENSE file in the repo. This is a summary of the license, not legal advice.", "source": "https://huggingface.co/qvac/BotPsy-0.8B/blob/main/LICENSE", "source_label": "LICENSE", "checked": "2026-11-11" }, { "kind": "official", "label": "Weight files", "text": "One BF16 safetensors file of 1,706,030,528 bytes (1.71 GB) at commit 24561e9681643eaf807fc28b0875210cb17f89c1. Hugging Face's safetensors total is 852,985,920 parameters. This is a download, not the memory a runtime needs.", "source": "https://huggingface.co/api/models/qvac/BotPsy-0.8B?blobs=true", "source_label": "Hugging Face API", "checked": "2026-11-11" }, { "kind": "official", "label": "GGUF files", "text": "Tether's GGUF repo (commit 17bb36aa29c58fcde8e321aaaad3f2f7372aa655) uses the qwen35 architecture and has an imatrix-calibrated Q4_K_M, a Q8_0 and a BF16 file. The card says the base model's MTP layers are not included. Every size is in the table below.", "source": "https://huggingface.co/qvac/BotPsy-0.8B-GGUFs", "source_label": "GGUF repo", "checked": "2026-11-11" }, { "kind": "vendor", "label": "Quant advice", "text": "Tether's GGUF card calls Q8_0 \"Recommended: scores like BF16 at about half the size\" and says Q4_K_M is \"about 1.5 points lower on average\". Those are Tether's claims, not checked by us. The fit on this page uses the Q4_K_M file.", "source": "https://huggingface.co/qvac/BotPsy-0.8B-GGUFs", "source_label": "GGUF card", "checked": "2026-11-11" }, { "kind": "vendor", "label": "Benchmark scores", "text": "Tether's card and blog report tool-calling and instruction-following scores in thinking mode against the base model. Those are Tether's own reported scores, not checked by us, and this page does not list them.", "source": "https://huggingface.co/blog/qvac/botpsy", "source_label": "Tether blog", "checked": "2026-11-11" }, { "kind": "checked", "label": "MLX", "text": "No MLX build of BotPsy-0.8B on Hugging Face as of 2026-11-11.", "source": "https://huggingface.co/models?search=BotPsy&library=mlx", "source_label": "Hugging Face MLX search", "checked": "2026-11-11" } ], "kv": { "layers": 6, "kv_heads": 2, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/qvac/BotPsy-0.8B/blob/main/config.json" } }, { "id": "swift-1.5-qwen3.8-flash-next", "name": "Swift 1.5 Qwen3.8-Flash-Next", "repo": "ukisai/Swift-Qwen3.8-Flash-Next", "params_b": 180.0, "params_source": "https://huggingface.co/ukisai/Swift-Qwen3.8-Flash-Next", "active_params_b": 6, "active_source": "https://huggingface.co/Qwen/Qwen3.8-Flash-Next", "q4_file_gb": 119.6, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-Flash-Next-GGUF", "accessed": "2026-11-11", "license": "Swift Open License v1.0", "license_source": "https://huggingface.co/ukisai/Swift-Qwen3.8-Flash-Next/blob/main/LICENSE", "context": 262144, "released": "2026-09-22", "released_source": "https://huggingface.co/ukisai/Swift-Qwen3.8-Flash-Next", "note": "UkisAI's reasoning-efficient fine-tune of Qwen3.8-Flash-Next: shorter thinking traces, tuned for coding and long agent tasks. Its config.json is identical to the base model's, so 6B active is Qwen's published count for the base; UkisAI's card does not restate it. UkisAI claims \"63.4% fewer thinking tokens, with a 1.8x speed up while keeping the accuracy loss <1% vs base on xhigh\"; the 1.8x is fewer tokens to finish one demo prompt, not faster decoding (details below). The checkpoint keeps the base model's 1-layer MTP head, so MTP self-speculative decoding is possible; that is not a speed figure. Q4_K_M is UkisAI's official GGUF (3 parts, 119,599,029,952 bytes); the vision projector is extra. License: free for commercial use only up to US$1,000,000 gross annual revenue (with affiliates), and the Qwen Community License 1.0 also applies (details below). Release date is the Hugging Face repo creation date.", "quants": [ { "quant": "Q4_K_M", "bytes": 119599029952, "source": "https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-Flash-Next-GGUF", "by": "UkisAI", "file": "Q4_K_M/Swift-1.5-Qwen3.8-Flash-Next-Q4_K_M (3 parts)" }, { "quant": "GSQ-RCO IQ3_S (mixed precision)", "bytes": 83743737120, "source": "https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF", "by": "UkisAI", "file": "Swift-Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S (2 parts)" }, { "quant": "GSQ-RCO IQ3_XXS (mixed precision)", "bytes": 75966073120, "source": "https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF", "by": "UkisAI", "file": "Swift-Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS (2 parts)" }, { "quant": "GSQ-RCO IQ2_XS (mixed precision)", "bytes": 68152167168, "source": "https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF", "by": "UkisAI", "file": "Swift-Qwen3.8-Flash-Next-GSQ-RCO-IQ2_XS (2 parts)" }, { "quant": "GSQ-RCO Q2_0 (mixed precision, experimental)", "bytes": 66549952800, "source": "https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF", "by": "UkisAI", "file": "Swift-Qwen3.8-Flash-Next-GSQ-RCO-Q2_0 (2 parts)" }, { "quant": "mmproj-BF16", "bytes": 907543520, "source": "https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF", "by": "UkisAI", "file": "mmproj-Swift-Qwen3.8-Flash-Next-BF16.gguf", "projector": true } ], "weights": { "bytes": 360000192888, "shards": 131, "revision": "0bd4fe22431372cdad1979267d3ab45aa7e6150a", "source": "https://huggingface.co/api/models/ukisai/Swift-Qwen3.8-Flash-Next?blobs=true", "note": "BF16 safetensors. Commit 0bd4fe22431372cdad1979267d3ab45aa7e6150a." }, "checks": [ { "kind": "official", "label": "License", "text": "UkisAI's contribution is under the Swift Open License v1.0. The card says it is free for personal, research, educational, evaluation and commercial use for individuals and organizations with gross annual revenue (including affiliates) up to US$1,000,000; above that, commercial use needs a separate Swift Enterprise License. This is a summary of the license, not legal advice.", "source": "https://huggingface.co/ukisai/Swift-Qwen3.8-Flash-Next/blob/main/LICENSE", "source_label": "LICENSE", "checked": "2026-11-11" }, { "kind": "official", "label": "Base model license", "text": "The base model's Qwen Community License 1.0 still applies. Under it, organizations running a Model-as-a-Service or AI Work Assistant business need a separate license from Qwen before commercial use, and products above 100 million monthly active users or US$20 million monthly revenue must display the model name. This is a summary of the license, not legal advice.", "source": "https://huggingface.co/ukisai/Swift-Qwen3.8-Flash-Next/blob/main/LICENSE-QWEN", "source_label": "LICENSE-QWEN", "checked": "2026-11-11" }, { "kind": "official", "label": "Architecture", "text": "config.json is identical to Qwen/Qwen3.8-Flash-Next's: model_type qwen4_exp, 48 layers (12 full attention), 2 KV heads, head size 256, 512 experts with 10 routed per token, 1 MTP layer and a vision tower. Hugging Face's safetensors total is 179,999,981,424 parameters, the same as the base. 6B active is Qwen's published figure for the base model. The card mentions YaRN scaling to 1,000,000 tokens, following Qwen's card; this page uses the config's 262,144.", "source": "https://huggingface.co/ukisai/Swift-Qwen3.8-Flash-Next/blob/main/config.json", "source_label": "config.json", "checked": "2026-11-11" }, { "kind": "official", "label": "Weight files", "text": "131 BF16 safetensors shards sum to 360,000,192,888 bytes (360.00 GB) at commit 0bd4fe22431372cdad1979267d3ab45aa7e6150a. UkisAI: \"The BF16 checkpoint is about 360 GB, so it needs a multi-GPU node. We evaluated it with tensor parallelism 8.\" These are downloads, not the memory a runtime needs. The repo was renamed from ukisai/Swift1.5-Qwen3.8-Flash-Next, which redirects.", "source": "https://huggingface.co/api/models/ukisai/Swift-Qwen3.8-Flash-Next?blobs=true", "source_label": "Hugging Face API", "checked": "2026-11-11" }, { "kind": "official", "label": "GGUF files", "text": "UkisAI's standard GGUF repo (commit 0142554c963a8653912c0c8283f05f13f63653f3, architecture qwen4exp) has the Q4_K_M in 3 parts: 44,648,342,368 + 44,928,589,600 + 30,022,097,984 bytes. A second repo, Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF (commit 99bb8f7f95c7aa7b24a36a7786a4f657b30f5d3d, created 2026-09-24, last modified 2026-10-07), has compact GSQ-RCO files. UkisAI says the GSQ-RCO tier names describe mixed-precision allocation profiles, not plain llama.cpp quants. It labels Q2_0 experimental and says \"For a file of similar size, prefer IQ2_XS\". Every size is in the table below.", "source": "https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-Flash-Next-GGUF", "source_label": "GGUF repo", "checked": "2026-11-11" }, { "kind": "vendor", "label": "UkisAI's efficiency claim", "text": "UkisAI's card: \"63.4% fewer thinking tokens, with a 1.8x speed up while keeping the accuracy loss <1% vs base on xhigh\". The 63.4% is the median thinking-token cut on GPQA-Diamond at reasoning effort xhigh (the mean cut there is 55.8%); at medium and low effort the score drops 2.62 and 2.42 points. The 1.8x comes from one demo prompt: 8 min 52 s for the base and 4 min 56 s for Swift to build a game. That is fewer tokens to finish, not faster decoding. UkisAI measured its scores in BF16 with MTP disabled. These are UkisAI's claims, not checked by us.", "source": "https://huggingface.co/ukisai/Swift-Qwen3.8-Flash-Next", "source_label": "Model card", "checked": "2026-11-11" }, { "kind": "vendor", "label": "MTP drafting", "text": "The checkpoint keeps the base model's 1-layer MTP head, and the card gives vLLM and SGLang flags for MTP self-speculative decoding. That makes drafting possible; it is not a speed figure, and the speed estimates on this page do not assume it.", "source": "https://huggingface.co/ukisai/Swift-Qwen3.8-Flash-Next", "source_label": "Model card", "checked": "2026-11-11" }, { "kind": "community", "label": "bartowski GGUF", "text": "bartowski/ukisai_Swift-Qwen3.8-Flash-Next-GGUF also exists. We did not size it; this page uses UkisAI's own files.", "source": "https://huggingface.co/bartowski/ukisai_Swift-Qwen3.8-Flash-Next-GGUF", "source_label": "bartowski repo", "checked": "2026-11-11" }, { "kind": "checked", "label": "MLX (Mac)", "text": "There is no official UkisAI MLX build of Swift Flash Next. UkisAI's Swift-1.5-4bit-MLX is the 27B model, not this one. The community builds in the next two rows are unchecked by us.", "source": "https://huggingface.co/ukisai", "source_label": "UkisAI on Hugging Face", "checked": "2026-11-11" }, { "kind": "community", "label": "Dankpaws MLX build", "text": "Dankpaws/Swift1.5-Qwen3.8-Flash-Next-MLX-4.7bpw is 106,723,353,968 bytes in total. Its card (the uploader's claim) says it targets a 96 GB Mac Studio with M5 Ultra, runs on mlx-serve 26.10.1 rather than stock mlx-lm, is text only, and keeps 32.0 GB of that as an n-gram table memory-mapped on disk (74.7 GB of weight shards). We have not run it and do not list its speed table; Mac fits on this page come only from our estimate for the Q4_K_M file.", "source": "https://huggingface.co/Dankpaws/Swift1.5-Qwen3.8-Flash-Next-MLX-4.7bpw", "source_label": "Dankpaws card", "checked": "2026-11-11" }, { "kind": "community", "label": "oMLX oQ builds", "text": "arkham00 publishes oMLX \"oQ\" mixed-precision builds, such as Swift-Qwen3.8-Flash-Next-oQ3.5e-mtp (89,777,274,008 bytes) and Swift-Qwen3.8-Flash-Next-oQ4e-mtp (106,275,556,357 bytes). Those are the uploader's builds. We have not run them and do not list their speed tables.", "source": "https://huggingface.co/arkham00/Swift-Qwen3.8-Flash-Next-oQ3.5e-mtp", "source_label": "arkham00 card", "checked": "2026-11-11" } ], "related": [ { "href": "/models/qwen3.8-flash-next/", "k": "Model", "t": "Base model: Qwen3.8-Flash-Next" } ], "kv": { "layers": 12, "kv_heads": 2, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/ukisai/Swift-Qwen3.8-Flash-Next/blob/main/config.json" }, "moe_skip": "derived active parameters (61.56B, 60.29B without embeddings) are more than 25% from the published 6B", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/ukisai/Swift-Qwen3.8-Flash-Next", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" } ] }, { "id": "raon-1-nano", "name": "Raon-1 nano", "repo": "KRAFTON/Raon-1-nano", "params_b": 3.1, "params_source": "https://huggingface.co/KRAFTON/Raon-1-nano", "active_params_b": null, "active_source": null, "q4_file_gb": 2.08, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/KRAFTON/Raon-1-nano-gguf", "accessed": "2026-11-11", "license": "Apache 2.0", "license_source": "https://huggingface.co/KRAFTON/Raon-1-nano", "context": 262144, "released": "2026-11-11", "released_source": "https://x.com/Krafton_AI/status/2120520399322861960", "note": "KRAFTON AI's small model for chat, instruction following and tool use. KRAFTON derived it from Raon-1 mini preview (30B-A4B) by MoE-to-Dense pruning and distillation, then trained it further on 3T tokens, followed by SFT, OPD and RL. It is dense, and a hybrid: of its 24 layers, 18 are Raon Delta Attention (linear attention with a fixed-size state) and 6 are gated MLA (multi-head latent attention). Thinking is on by default, and KRAFTON says turning it off is not recommended. KRAFTON gives 3.1B parameters, including a 1.5B per-layer embedding table, and calls it an effective 1.5B. Those per-layer embeddings still sit in memory, so the fit counts the whole file. KRAFTON's other cards say 3.2B (details below). Q4_K_M is KRAFTON's official GGUF (raon-1-nano-Q4_K_M.gguf, 2,082,677,632 bytes), made with KRAFTON's own llama.cpp fork; every size is in the table below. The _MTP files add a small MTP head for speculative decoding (drafting tokens ahead); the fit and the speed estimates use the plain file and plain decoding. Apache 2.0, including commercial use of the downloaded weights; KRAFTON's hosted API and playground have separate terms. Release date is KRAFTON's announcement; the Hugging Face repo was created on Oct 29, 2026.", "runtime_support": { "status": "KRAFTON's llama.cpp fork and MLX build", "text": "Stock llama.cpp, Ollama and LM Studio can't load Raon-1 nano yet. Upstream llama.cpp has no Raon-1 architecture (these GGUFs use raon1nano), and we found no upstream pull request for nano. KRAFTON's GGUF card builds from its fork: git clone -b raon-1-nano-preview https://github.com/krafton-ai/llama.cpp (the card also mentions a raon-1-nano-optimize branch). On a Mac, KRAFTON's MLX build (KRAFTON/Raon-1-nano-mlx) needs trust_remote_code=True and pins mlx==0.32.4, mlx-lm==0.32.2 and transformers==5.20.2. The fits and speeds on this page are math on the official Q4_K_M file, not a test on the fork.", "source": "https://huggingface.co/KRAFTON/Raon-1-nano-gguf", "checked": "2026-11-11" }, "quants": [ { "quant": "Q4_K_M", "bytes": 2082677632, "source": "https://huggingface.co/KRAFTON/Raon-1-nano-gguf", "by": "KRAFTON", "file": "raon-1-nano-Q4_K_M.gguf" }, { "quant": "Q4_K_M + MTP head", "bytes": 2145934208, "source": "https://huggingface.co/KRAFTON/Raon-1-nano-gguf", "by": "KRAFTON", "file": "raon-1-nano-Q4_K_M_MTP.gguf" }, { "quant": "Q6_K", "bytes": 2646016384, "source": "https://huggingface.co/KRAFTON/Raon-1-nano-gguf", "by": "KRAFTON", "file": "raon-1-nano-Q6_K.gguf" }, { "quant": "Q6_K + MTP head", "bytes": 2709272960, "source": "https://huggingface.co/KRAFTON/Raon-1-nano-gguf", "by": "KRAFTON", "file": "raon-1-nano-Q6_K_MTP.gguf" }, { "quant": "Q8_0", "bytes": 3391016320, "source": "https://huggingface.co/KRAFTON/Raon-1-nano-gguf", "by": "KRAFTON", "file": "raon-1-nano-Q8_0.gguf" }, { "quant": "Q8_0 + MTP head", "bytes": 3454272896, "source": "https://huggingface.co/KRAFTON/Raon-1-nano-gguf", "by": "KRAFTON", "file": "raon-1-nano-Q8_0_MTP.gguf" }, { "quant": "F16", "bytes": 6274887040, "source": "https://huggingface.co/KRAFTON/Raon-1-nano-gguf", "by": "KRAFTON", "file": "raon-1-nano-F16.gguf" }, { "quant": "F16 + MTP head", "bytes": 6384592256, "source": "https://huggingface.co/KRAFTON/Raon-1-nano-gguf", "by": "KRAFTON", "file": "raon-1-nano-F16_MTP.gguf" } ], "weights": { "bytes": 7060010384, "shards": 3, "source": "https://huggingface.co/api/models/KRAFTON/Raon-1-nano?blobs=true", "note": "BF16 safetensors: two model shards plus model-tied-lm-head.safetensors (798,228,608 bytes). The 55M MTP module is a separate file (mtp/model.safetensors, 109,687,880 bytes)." }, "checks": [ { "kind": "official", "label": "Model card", "text": "KRAFTON's card: a dense hybrid with 24 layers (18 Raon Delta Attention, 6 gated MLA), hidden size 1,536, SwiGLU, Raon-mHC with 4 residual streams, no position encoding (NoPE), a 259,840-token vocabulary, a tied output head and a per-layer embedding size of 256. Native context is 262,144 tokens. Thinking is on by default. The checkpoint also has a 1-layer, 55M-parameter MTP module.", "source": "https://huggingface.co/KRAFTON/Raon-1-nano", "source_label": "Model card", "checked": "2026-11-11" }, { "kind": "official", "label": "Parameter count", "text": "KRAFTON's announcement and main card say 3.1B in total, including a 1.5B per-layer embedding table (an effective 1.5B). The main card also says the Hugging Face checkpoint is 3.2B with the 55M MTP module. The GGUF and MLX cards say 3.2B, including a 1.6B table. Hugging Face's own counter reads 3,529,911,376, which may count the tied output head twice (not checked). This page uses KRAFTON's 3.1B. The model is dense, so no active count is used. Its per-layer embedding table is a lookup, so each token reads only a small slice of it, but the table still sits in memory. The speed estimate on this page reads the whole Q4_K_M file per token, so it may be on the slow side.", "source": "https://huggingface.co/KRAFTON/Raon-1-nano", "source_label": "Model card", "checked": "2026-11-11" }, { "kind": "official", "label": "License", "text": "Apache 2.0, including commercial use of the downloaded weights, per the cards. KRAFTON's hosted API and playground have separate terms. The repos are not gated. This is a summary of the license, not legal advice.", "source": "https://huggingface.co/KRAFTON/Raon-1-nano", "source_label": "Model card", "checked": "2026-11-11" }, { "kind": "official", "label": "Weight files", "text": "Three BF16 safetensors files sum to 7,060,010,384 bytes (7.06 GB): two model shards (3,068,867,712 and 3,192,914,064 bytes) and model-tied-lm-head.safetensors (798,228,608 bytes). The MTP module is a separate 109,687,880-byte file. These are downloads, not the memory a runtime needs.", "source": "https://huggingface.co/api/models/KRAFTON/Raon-1-nano?blobs=true", "source_label": "Hugging Face API", "checked": "2026-11-11" }, { "kind": "official", "label": "GGUF files", "text": "KRAFTON's own GGUF repo, made with the raon-1-nano-preview branch of its llama.cpp fork (GGUF architecture raon1nano). It has Q4_K_M, Q6_K, Q8_0 and F16, each also as an _MTP file with the MTP head. The fit on this page uses the plain Q4_K_M. We found no third-party GGUFs. Every size is in the table below.", "source": "https://huggingface.co/KRAFTON/Raon-1-nano-gguf", "source_label": "GGUF repo", "checked": "2026-11-11" }, { "kind": "official", "label": "MLX build", "text": "KRAFTON/Raon-1-nano-mlx is 4-bit (GPTQ, group size 64): model.safetensors is 1,805,577,858 bytes, plus a 62,411,311-byte MTP file. It ships its own model code (raon_mlx_model.py), so it needs trust_remote_code=True, and it pins mlx==0.32.4, mlx-lm==0.32.2 and transformers==5.20.2. KRAFTON says its fused Metal kernels \"are enabled only with mlx==0.32.4 (other versions run a slower path)\". KRAFTON puts a BF16 MLX export at about 6.3 GB and says its decode is slower than 4-bit.", "source": "https://huggingface.co/KRAFTON/Raon-1-nano-mlx", "source_label": "MLX repo", "checked": "2026-11-11" }, { "kind": "checked", "label": "Upstream llama.cpp", "text": "ggml-org/llama.cpp master has no Raon-1 model file (src/models/raon1.cpp is a 404), and we found no upstream pull request for Raon-1 nano. The only related one, #31107 for Raon-1 mini preview, was closed unmerged on Oct 24, 2026.", "source": "https://github.com/ggml-org/llama.cpp/pull/31107", "source_label": "llama.cpp PR #31107", "checked": "2026-11-11" }, { "kind": "checked", "label": "Ollama and LM Studio", "text": "An Ollama search on 2026-11-11 found no KRAFTON Raon model, only unrelated names. We found no LM Studio listing.", "source": "https://ollama.com/search?q=raon", "source_label": "Ollama search", "checked": "2026-11-11" }, { "kind": "vendor", "label": "Other runtimes", "text": "KRAFTON also publishes a LiteRT-LM build (KRAFTON/Raon-1-nano-litert-lm). Its card says it runs with a 4,096-token context and answers without thinking, and that the first GPU compile takes about 30 to 60 seconds. Transformers needs trust_remote_code; KRAFTON notes that Raon Delta Attention prefill runs one step at a time, so long prompts can be slow to process. KRAFTON's vLLM instructions install its own fork (github.com/krafton-ai/vllm, branch raon-1-nano). We did not check upstream vLLM.", "source": "https://huggingface.co/KRAFTON/Raon-1-nano", "source_label": "Model card", "checked": "2026-11-11" }, { "kind": "vendor", "label": "Benchmark scores (KRAFTON's)", "text": "KRAFTON's card compares nano with Qwen3.5 2B, Gemma 4 E2B, MiniCPM5 2B and LFM3 1.7B. Its scores for nano: MMLU-Redux 81.1, IFBench 75.9, MultiChallenge 40.5, FACTS Grounding 44.4, BFCL v4 56.3, BrowseComp Plus 34.5, NL2SH-ALFA 58.0, AndroidControl 36.0 (average of HL and LL), PersonaMem v2 38.8 (32K setting) and PrefEval 44.0. In Korean: KMMLU-Redux 50.9, HAE-RAE 1.1 44.2 and CLIcK 65.2. Office Work Bench (49.2) and Chatbot Korean (91.0) are KRAFTON's internal tests and are not public. Sampling settings differ by benchmark and are on the card. These are KRAFTON's scores, not checked by us. KRAFTON publishes no speed figures for nano, so the speed estimates on this page are our own formula for plain decoding with the Q4_K_M file, without MTP drafting.", "source": "https://huggingface.co/KRAFTON/Raon-1-nano", "source_label": "Model card", "checked": "2026-11-11" } ], "related": [ { "href": "/models/raon-1-mini-preview/", "k": "Model", "t": "Raon-1 mini preview, the larger MoE model" }, { "href": "/models/lfm3-1.7b/", "k": "Model", "t": "LFM3-1.7B, one of the models KRAFTON compares against" }, { "href": "/models/gemma-4-e2b-it/", "k": "Model", "t": "Gemma 4 E2B, another per-layer embedding model" } ], "kv_skip": "Raon-1 nano's architecture is not in upstream llama.cpp, so there is no verified llama.cpp KV formula for it yet. KRAFTON says only its 6 gated MLA layers keep a KV cache, about 7 KiB per token in BF16, and the 18 Raon Delta Attention layers keep a fixed-size state. That is KRAFTON's figure, not checked here", "kv": { "arch": "hybrid" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/KRAFTON/Raon-1-nano", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" } ] }, { "id": "raon-1-mini-preview", "name": "Raon-1 mini preview", "repo": "KRAFTON/Raon-1-mini-preview", "params_b": 28.8, "params_source": "https://huggingface.co/KRAFTON/Raon-1-mini-preview", "active_params_b": 4.3, "active_source": "https://huggingface.co/KRAFTON/Raon-1-mini-preview", "q4_file_gb": 19.69, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/KRAFTON/Raon-1-mini-preview-gguf", "accessed": "2026-11-11", "license": "Apache 2.0", "license_source": "https://huggingface.co/KRAFTON/Raon-1-mini-preview", "context": 262144, "released": "2026-10-21", "released_source": "https://huggingface.co/KRAFTON/Raon-1-mini-preview", "note": "KRAFTON AI's mixture-of-experts model, trained from scratch, for reasoning, coding and agent work with tool calling. It is a preview: KRAFTON says specs and results may change before the final version. Of its 25 layers, 18 are Raon Delta Attention (linear attention with a fixed-size state) and 7 are gated MLA (multi-head latent attention). It has 128 routed experts, and each token uses 6 of them plus 1 shared expert. KRAFTON gives 28.8B total and 4.3B active parameters. Q4_K_M is KRAFTON's official GGUF (raon-1-mini-preview-Q4_K_M.gguf, 19,686,434,240 bytes), made with KRAFTON's own llama.cpp fork; a smaller official IQ4_XS (16.07 GB) and every other size are in the table below. Before downloading, you request access once on the Hugging Face page (KRAFTON says approval is automatic) and log in. Apache 2.0, including commercial use of the downloaded weights; KRAFTON's hosted API and playground have separate terms. Release date is the Hugging Face repo creation date (Oct 21, 2026, UTC).", "runtime_support": { "status": "KRAFTON's llama.cpp fork and MLX build", "text": "Stock llama.cpp, Ollama and LM Studio can't load Raon-1 mini preview yet. Upstream llama.cpp has no Raon-1 architecture (these GGUFs use raon1). Pull request #31107, which would have added it, was closed unmerged on Oct 24, 2026, after a maintainer asked KRAFTON to submit its own code. KRAFTON's GGUF card says to build from its fork: git clone -b raon-1-mini-preview https://github.com/krafton-ai/llama.cpp. On a Mac, KRAFTON's MLX build (KRAFTON/Raon-1-mini-preview-mlx, in NVIDIA's 4-bit NVFP4 format) needs trust_remote_code=True and pins mlx==0.32.4, mlx-lm==0.32.2 and transformers==5.20.1. The fits and speeds on this page are math on the official Q4_K_M file, not a test on the fork.", "source": "https://huggingface.co/KRAFTON/Raon-1-mini-preview-gguf", "checked": "2026-11-11" }, "quants": [ { "quant": "IQ4_XS", "bytes": 16071801280, "source": "https://huggingface.co/KRAFTON/Raon-1-mini-preview-gguf", "by": "KRAFTON", "file": "raon-1-mini-preview-IQ4_XS.gguf" }, { "quant": "Q4_K_M", "bytes": 19686434240, "source": "https://huggingface.co/KRAFTON/Raon-1-mini-preview-gguf", "by": "KRAFTON", "file": "raon-1-mini-preview-Q4_K_M.gguf" }, { "quant": "Q5_K_M", "bytes": 22183765440, "source": "https://huggingface.co/KRAFTON/Raon-1-mini-preview-gguf", "by": "KRAFTON", "file": "raon-1-mini-preview-Q5_K_M.gguf" }, { "quant": "Q6_K", "bytes": 25839462080, "source": "https://huggingface.co/KRAFTON/Raon-1-mini-preview-gguf", "by": "KRAFTON", "file": "raon-1-mini-preview-Q6_K.gguf" }, { "quant": "Q8_0", "bytes": 30693555936, "source": "https://huggingface.co/KRAFTON/Raon-1-mini-preview-gguf", "by": "KRAFTON", "file": "raon-1-mini-preview-Q8_0.gguf" } ], "weights": { "bytes": 32702145088, "shards": 7, "source": "https://huggingface.co/api/models/KRAFTON/Raon-1-mini-preview?blobs=true", "note": "Seven safetensors shards: BF16, with the routed experts stored in MXFP8." }, "checks": [ { "kind": "official", "label": "Model card", "text": "KRAFTON's card: a 30B-A4B mixture-of-experts model trained from scratch. 25 layers (18 Raon Delta Attention, 7 gated MLA), hidden size 2,560, SiTU-GLU, Raon-mHC with 4 residual streams, no position encoding (NoPE), a 259,840-token vocabulary and an untied output head. 128 routed experts (expert intermediate size 1,088), 6 routed and 1 shared active per token; the first layer is a dense FFN. Native context is 262,144 tokens, with thinking built in. The card covers tool calling with Hermes Agent, Pi and OpenCode, and lists Korean, English, Japanese, Chinese, Spanish, French and Thai. It says: \"This is a preview release, not a final model.\"", "source": "https://huggingface.co/KRAFTON/Raon-1-mini-preview", "source_label": "Model card", "checked": "2026-11-11" }, { "kind": "official", "label": "Parameters and config", "text": "KRAFTON's table: 28.8B total, 4.3B activated. Hugging Face's counter reads 28,783,218,402 (3,113,981,970 BF16, 25,669,140,480 F8_E4M3 and 95,952 F32), because the main repo stores the routed experts in MXFP8. config.json: 25 layers, hidden size 2,560, 128 experts, 6 per token, 1 shared, q_lora_rank 576, kv_lora_rank 512, untied embeddings, 262,144 max positions.", "source": "https://huggingface.co/KRAFTON/Raon-1-mini-preview/blob/main/config.json", "source_label": "config.json", "checked": "2026-11-11" }, { "kind": "official", "label": "Access and license", "text": "Apache 2.0, including commercial use of the downloaded weights. The hosted API and playground have separate terms. KRAFTON's cards say: \"The model repositories are gated. Request access once on this page (approval is automatic; it covers every Raon-1 mini preview repository), then log in so the examples below can download them.\" Every card shows that request form, though Hugging Face's API reports the repos as not gated. Expect the request step either way. This is a summary of the license, not legal advice.", "source": "https://huggingface.co/KRAFTON/Raon-1-mini-preview", "source_label": "Model card", "checked": "2026-11-11" }, { "kind": "official", "label": "Weight files", "text": "Seven safetensors shards sum to 32,702,145,088 bytes (32.70 GB); the index lists 32,699,649,368 bytes of tensors. These are downloads, not the memory a runtime needs.", "source": "https://huggingface.co/api/models/KRAFTON/Raon-1-mini-preview?blobs=true", "source_label": "Hugging Face API", "checked": "2026-11-11" }, { "kind": "official", "label": "GGUF files", "text": "KRAFTON's own GGUF repo (GGUF architecture raon1). The card says the files were converted from the BF16 weights of the same checkpoint, and that Q6_K and smaller used an importance matrix (imatrix). It has IQ4_XS, Q4_K_M, Q5_K_M, Q6_K and Q8_0; the fit on this page uses Q4_K_M. There is also a 732,539,232-byte DFlash2 draft model (dflash-raon-1-mini-preview-Q8_0.gguf), which is only for speculative decoding and is not a fit file. Every size is in the table below.", "source": "https://huggingface.co/KRAFTON/Raon-1-mini-preview-gguf", "source_label": "GGUF repo", "checked": "2026-11-11" }, { "kind": "official", "label": "MLX build", "text": "KRAFTON/Raon-1-mini-preview-mlx is NVFP4, in four shards summing to 17,537,518,992 bytes, plus a 207,872,547-byte head file and a 971,364,975-byte DFlash2 draft. It ships its own model code, so it needs trust_remote_code=True, and it pins mlx==0.32.4, mlx-lm==0.32.2 and transformers==5.20.1. KRAFTON says the DFlash2 draft needs mlx 0.32.4 and mlx-lm 0.32.x (other versions warn and run without it) and adds about 1 GB of memory.", "source": "https://huggingface.co/KRAFTON/Raon-1-mini-preview-mlx", "source_label": "MLX repo", "checked": "2026-11-11" }, { "kind": "vendor", "label": "NVFP4 on an RTX 5090 (KRAFTON's claim)", "text": "KRAFTON's card says the NVFP4 format enables inference on a single NVIDIA RTX 5090 (32 GB). For vLLM it points to a separate checkpoint, KRAFTON/Raon-1-mini-preview-nvfp4. That is KRAFTON's claim, not a fit we checked. The fits on this page use the Q4_K_M GGUF.", "source": "https://huggingface.co/KRAFTON/Raon-1-mini-preview", "source_label": "Model card", "checked": "2026-11-11" }, { "kind": "vendor", "label": "DFlash2 speculative decoding (KRAFTON's figures)", "text": "KRAFTON says its DFlash2 draft model gives up to 2.18× the output throughput of plain decoding on a single NVIDIA B300 (GSM8K, one request at a time, vLLM 0.28.0 with the MXFP8 model). Its card adds that speedups vary by hardware and workload. This is speculative decoding on a datacenter GPU, not plain decoding. The speed estimates on this page do not use it.", "source": "https://huggingface.co/KRAFTON/Raon-1-mini-preview", "source_label": "Model card", "checked": "2026-11-11" }, { "kind": "vendor", "label": "Benchmark scores (KRAFTON's)", "text": "KRAFTON's card compares the model with Qwen3.6-35B-A3B and Gemma 4 26B-A4B, with Qwen3.8-27B and Muse Glimmer 30B as dense references; gpt-5.6-luna was the judge and user simulator. Its scores for Raon-1 mini preview include MMLU Pro 76.5, GPQA Diamond 84.3, AIME 2026 93.3, IFBench 84.2, LiveCodeBench v6 81.6, SWE-Verified 70.8 (mini-SWE-agent harness), Terminal-Bench Lite 47.0, BFCL v4 59.8, BrowseComp Plus 55.7 and AA-LCR 67.3. Chatbot Korean (92.9) is KRAFTON's internal test and is not public. These are KRAFTON's scores, not checked by us.", "source": "https://huggingface.co/KRAFTON/Raon-1-mini-preview", "source_label": "Model card", "checked": "2026-11-11" }, { "kind": "checked", "label": "Upstream llama.cpp", "text": "Pull request #31107 (\"model : add Raon-1 mini preview support\") was closed without merging on Oct 24, 2026, after a maintainer asked KRAFTON to submit its own code. The matching request, issue #31106, is also closed. ggml-org/llama.cpp master has no src/models/raon1.cpp.", "source": "https://github.com/ggml-org/llama.cpp/pull/31107", "source_label": "llama.cpp PR #31107", "checked": "2026-11-11" }, { "kind": "checked", "label": "Ollama and LM Studio", "text": "An Ollama search on 2026-11-11 found no KRAFTON Raon model, only unrelated names. We found no LM Studio listing.", "source": "https://ollama.com/search?q=raon", "source_label": "Ollama search", "checked": "2026-11-11" } ], "related": [ { "href": "/models/raon-1-nano/", "k": "Model", "t": "Raon-1 nano, the small dense model" }, { "href": "/models/qwen3.6-35b-a3b/", "k": "Model", "t": "Qwen3.6-35B-A3B, one of the models KRAFTON compares against" }, { "href": "/models/gemma-4-26b-a4b-it/", "k": "Model", "t": "Gemma 4 26B A4B, another model KRAFTON compares against" } ], "kv_skip": "Raon-1 mini preview's architecture is not in upstream llama.cpp, so there is no verified llama.cpp KV formula for it yet. KRAFTON says only its 7 gated MLA layers keep a KV cache, about 8 KiB per token in BF16, and the 18 Raon Delta Attention layers keep a fixed-size state. That is KRAFTON's figure, not checked here", "kv": { "arch": "hybrid" }, "moe": { "expert_b": 25.669, "expert_active_b": 1.203, "layers": 24, "experts": 128, "experts_per_tok": 6, "expert_width": 1088, "config_source": "https://huggingface.co/KRAFTON/Raon-1-mini-preview/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/KRAFTON/Raon-1-mini-preview", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" } ] }, { "id": "plamo-3-translate-0.8b", "name": "PLaMo 3 Translate 0.8B", "repo": "pfnet/plamo-3-translate-0.8b", "params_b": 0.89, "params_source": "https://huggingface.co/pfnet/plamo-3-translate-0.8b", "active_params_b": null, "active_source": null, "q4_file_gb": 0.57, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/pfnet/plamo-3-translate-0.8b-GGUF", "accessed": "2026-11-10", "license": "PLaMo Community License", "license_source": "https://plamo.preferredai.jp/info/plamo-community-license-en", "context": 32768, "released": "2026-11-05", "released_source": "https://www.preferred.jp/en/blog/tech/plamo-3-translate-weight-release", "task": "translation", "note": "Translation model, not a general chat assistant: it takes text in one language and returns a translation, through its own translation prompt format. Dense model, one of four PLaMo 3 Translate sizes. Context is 32,768 tokens because the model card says training used at most 32,768 tokens of input and output combined and asks you to split anything longer; config.json allows 262,144 positions, which this page does not treat as usable. Commercial use needs advance registration with Preferred Networks (see License below). Release date is the date of PFN's release blog post.", "quants": [ { "quant": "Q4_K_M", "bytes": 573126848, "source": "https://huggingface.co/pfnet/plamo-3-translate-0.8b-GGUF", "by": "Preferred Networks" }, { "quant": "Q4_0", "bytes": 547543232, "source": "https://huggingface.co/pfnet/plamo-3-translate-0.8b-GGUF", "by": "Preferred Networks" }, { "quant": "Q3_K_M", "bytes": 486385856, "source": "https://huggingface.co/pfnet/plamo-3-translate-0.8b-GGUF", "by": "Preferred Networks" }, { "quant": "Q2_K", "bytes": 410138816, "source": "https://huggingface.co/pfnet/plamo-3-translate-0.8b-GGUF", "by": "Preferred Networks" } ], "related": [ { "href": "/models/plamo-3-translate-2b/", "k": "Model", "t": "PLaMo 3 Translate 2B, another size of the same translation model" }, { "href": "/models/plamo-3-translate-8b/", "k": "Model", "t": "PLaMo 3 Translate 8B, another size of the same translation model" }, { "href": "/models/plamo-3-translate-31b/", "k": "Model", "t": "PLaMo 3 Translate 31B, another size of the same translation model" }, { "href": "/models/", "k": "Catalog", "t": "Every model in the catalog" } ], "checks": [ { "kind": "official", "label": "Model card", "text": "PFN's card describes PLaMo 3 Translate as a translation-specialized model, tagged translation on Hugging Face, not a general chat or coding assistant. It uses a translation chat template (the PFML3 prompt format). The release blog covers four sizes (0.8B, 2B, 8B, 31B) and 53 languages, and says the model accepts surrounding context and glossaries and can write plain or polite Japanese. The card says the longest training sequence was 32,768 tokens for input and output combined and to split anything longer, so this page uses 32,768 as the context, not config.json's 262,144. Safetensors total 890,344,960 parameters (0.89B), BF16, dense. The repo is not gated.", "source": "https://huggingface.co/pfnet/plamo-3-translate-0.8b", "source_label": "Model card", "checked": "2026-11-10" }, { "kind": "official", "label": "GGUF files", "text": "PFN publishes the GGUF build itself (pfnet/plamo-3-translate-0.8b-GGUF), one file per quant: Q4_K_M 573,126,848 bytes, Q4_0 547,543,232 bytes, Q3_K_M 486,385,856 bytes, Q2_K 410,138,816 bytes. This page's fit and speed estimates use the Q4_K_M file. The 31B card's example runs llama-server with -c 16384, --jinja and --no-prefill-assistant; 16,384 is only their example. We read the sizes from the Hugging Face file list and did not download them.", "source": "https://huggingface.co/pfnet/plamo-3-translate-0.8b-GGUF", "source_label": "GGUF repo", "checked": "2026-11-10" }, { "kind": "official", "label": "MLX 4-bit build", "text": "PFN also publishes an MLX 4-bit build for Apple silicon (pfnet/plamo-3-translate-0.8b-MLX-4bit), 975,810,552 bytes of safetensors. Its card says to run it with mlx-lm with trust_remote_code=True. Other official formats in the release are BF16, FP8, W4A16, W4AFP8 and NVFP4. We have not run the MLX build on a Mac, and no MLX speed is listed.", "source": "https://huggingface.co/pfnet/plamo-3-translate-0.8b-MLX-4bit", "source_label": "MLX-4bit repo", "checked": "2026-11-10" }, { "kind": "official", "label": "License: commercial use needs registration", "text": "PLaMo Community License. Personal and non-commercial use, including academic use, is free with no registration. Commercial use is free only if you (with your affiliates) had revenue of JPY 1 billion or less in the most recent fiscal year AND you register with PFN in advance through its form (https://forms.gle/mTL8tBLrMYXKNZD56). Above JPY 1 billion you need a separate license agreement with PFN (https://translate.preferredai.jp/llm/contact/). This is a summary of the license, not legal advice.", "source": "https://plamo.preferredai.jp/info/plamo-community-license-en", "source_label": "PLaMo Community License", "checked": "2026-11-10" }, { "kind": "vendor", "label": "PFN's benchmark claims", "text": "PFN's own claim: even the 0.8B model outperforms existing translation-specialized models (TranslateGemma, Hy-MT2) on Japanese-English and on Japanese to and from other languages, and PFN calls PLaMo 3 Translate the world's best translation-specialized LLM. Its benchmarks are PFTB and SaaS (PFN internal sets), WMT24++ (COMET-22) and FLORES+ (chrF++). PFN's card also says the Q2_K GGUF loses accuracy on some metrics while the other formats stay close to BF16. These are PFN's results; we have not reproduced any of them.", "source": "https://www.preferred.jp/en/blog/tech/plamo-3-translate-weight-release", "source_label": "PFN release blog", "checked": "2026-11-10" }, { "kind": "checked", "label": "llama.cpp and mlx-lm support", "text": "llama.cpp master has a plamo3 architecture (src/models/plamo3.cpp), and the official GGUF header says general.architecture plamo3. With a sliding window set, llama.cpp loads a sliding-window pattern of 8: in each group of 8 layers, 7 use a 2,048-token sliding window and 1 uses full attention. For this size that is 2 full-attention and 14 sliding-window layers of 16, with 1 KV head of size 128, which is what the context estimates use. mlx-lm master has mlx_lm/models/plamo3.py.", "source": "https://github.com/ggml-org/llama.cpp/blob/master/src/models/plamo3.cpp", "source_label": "llama.cpp plamo3.cpp", "checked": "2026-11-10" }, { "kind": "unknown", "label": "Not tested by us", "text": "We have not run any PLaMo 3 Translate model. No speed on this page is measured: tok/s figures are the site's bandwidth estimates. Translation quality is not tested by us, and the sizes are not compared with each other or with other models here.", "source": "https://huggingface.co/pfnet/plamo-3-translate-0.8b", "source_label": "Model card does not give this", "checked": "2026-11-10" } ], "kv": { "layers": 2, "kv_heads": 1, "head_dim": 128, "v_dim": 128, "swa_layers": 14, "swa_kv_heads": 1, "swa_head_dim": 128, "window": 2048, "arch": "swa", "max_ctx": 32768, "config_source": "https://huggingface.co/pfnet/plamo-3-translate-0.8b/blob/main/config.json", "config_note": "16 layers, window_size 2048, sliding_window_pattern 8: llama.cpp plamo3 (load_swa_pattern(ml, 8), LLAMA_SWA_TYPE_STANDARD) makes layer il sliding-window unless il % 8 == 7, so 2 full-attention + 14 sliding-window layers; the official GGUF metadata says the same (sliding_window 2048, sliding_window_pattern 8). max_ctx is the model card's 32,768-token training limit for input and output combined (the context field), not config.json's max_position_embeddings 262,144" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/pfnet/plamo-3-translate-0.8b", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" } ] }, { "id": "plamo-3-translate-2b", "name": "PLaMo 3 Translate 2B", "repo": "pfnet/plamo-3-translate-2b", "params_b": 2.88, "params_source": "https://huggingface.co/pfnet/plamo-3-translate-2b", "active_params_b": null, "active_source": null, "q4_file_gb": 1.81, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/pfnet/plamo-3-translate-2b-GGUF", "accessed": "2026-11-10", "license": "PLaMo Community License", "license_source": "https://plamo.preferredai.jp/info/plamo-community-license-en", "context": 32768, "released": "2026-11-05", "released_source": "https://www.preferred.jp/en/blog/tech/plamo-3-translate-weight-release", "task": "translation", "note": "Translation model, not a general chat assistant: it takes text in one language and returns a translation, through its own translation prompt format. Dense model, one of four PLaMo 3 Translate sizes. Context is 32,768 tokens because the model card says training used at most 32,768 tokens of input and output combined and asks you to split anything longer; config.json allows 262,144 positions, which this page does not treat as usable. Commercial use needs advance registration with Preferred Networks (see License below). Release date is the date of PFN's release blog post.", "quants": [ { "quant": "Q4_K_M", "bytes": 1807203040, "source": "https://huggingface.co/pfnet/plamo-3-translate-2b-GGUF", "by": "Preferred Networks" }, { "quant": "Q4_0", "bytes": 1698577120, "source": "https://huggingface.co/pfnet/plamo-3-translate-2b-GGUF", "by": "Preferred Networks" }, { "quant": "Q3_K_M", "bytes": 1495886560, "source": "https://huggingface.co/pfnet/plamo-3-translate-2b-GGUF", "by": "Preferred Networks" }, { "quant": "Q2_K", "bytes": 1225612000, "source": "https://huggingface.co/pfnet/plamo-3-translate-2b-GGUF", "by": "Preferred Networks" } ], "related": [ { "href": "/models/plamo-3-translate-0.8b/", "k": "Model", "t": "PLaMo 3 Translate 0.8B, another size of the same translation model" }, { "href": "/models/plamo-3-translate-8b/", "k": "Model", "t": "PLaMo 3 Translate 8B, another size of the same translation model" }, { "href": "/models/plamo-3-translate-31b/", "k": "Model", "t": "PLaMo 3 Translate 31B, another size of the same translation model" }, { "href": "/models/", "k": "Catalog", "t": "Every model in the catalog" } ], "checks": [ { "kind": "official", "label": "Model card", "text": "PFN's card describes PLaMo 3 Translate as a translation-specialized model, tagged translation on Hugging Face, not a general chat or coding assistant. It uses a translation chat template (the PFML3 prompt format). The release blog covers four sizes (0.8B, 2B, 8B, 31B) and 53 languages, and says the model accepts surrounding context and glossaries and can write plain or polite Japanese. The card says the longest training sequence was 32,768 tokens for input and output combined and to split anything longer, so this page uses 32,768 as the context, not config.json's 262,144. Safetensors total 2,878,595,584 parameters (2.88B), BF16, dense. The repo is not gated.", "source": "https://huggingface.co/pfnet/plamo-3-translate-2b", "source_label": "Model card", "checked": "2026-11-10" }, { "kind": "official", "label": "GGUF files", "text": "PFN publishes the GGUF build itself (pfnet/plamo-3-translate-2b-GGUF), one file per quant: Q4_K_M 1,807,203,040 bytes, Q4_0 1,698,577,120 bytes, Q3_K_M 1,495,886,560 bytes, Q2_K 1,225,612,000 bytes. This page's fit and speed estimates use the Q4_K_M file. The 31B card's example runs llama-server with -c 16384, --jinja and --no-prefill-assistant; 16,384 is only their example. We read the sizes from the Hugging Face file list and did not download them.", "source": "https://huggingface.co/pfnet/plamo-3-translate-2b-GGUF", "source_label": "GGUF repo", "checked": "2026-11-10" }, { "kind": "official", "label": "MLX 4-bit build", "text": "PFN also publishes an MLX 4-bit build for Apple silicon (pfnet/plamo-3-translate-2b-MLX-4bit), 2,410,973,952 bytes of safetensors. Its card says to run it with mlx-lm with trust_remote_code=True. Other official formats in the release are BF16, FP8, W4A16, W4AFP8 and NVFP4. We have not run the MLX build on a Mac, and no MLX speed is listed.", "source": "https://huggingface.co/pfnet/plamo-3-translate-2b-MLX-4bit", "source_label": "MLX-4bit repo", "checked": "2026-11-10" }, { "kind": "official", "label": "License: commercial use needs registration", "text": "PLaMo Community License. Personal and non-commercial use, including academic use, is free with no registration. Commercial use is free only if you (with your affiliates) had revenue of JPY 1 billion or less in the most recent fiscal year AND you register with PFN in advance through its form (https://forms.gle/mTL8tBLrMYXKNZD56). Above JPY 1 billion you need a separate license agreement with PFN (https://translate.preferredai.jp/llm/contact/). This is a summary of the license, not legal advice.", "source": "https://plamo.preferredai.jp/info/plamo-community-license-en", "source_label": "PLaMo Community License", "checked": "2026-11-10" }, { "kind": "vendor", "label": "PFN's benchmark claims", "text": "PFN's own claim: even the 0.8B model outperforms existing translation-specialized models (TranslateGemma, Hy-MT2) on Japanese-English and on Japanese to and from other languages, and PFN calls PLaMo 3 Translate the world's best translation-specialized LLM. Its benchmarks are PFTB and SaaS (PFN internal sets), WMT24++ (COMET-22) and FLORES+ (chrF++). PFN's card also says the Q2_K GGUF loses accuracy on some metrics while the other formats stay close to BF16. These are PFN's results; we have not reproduced any of them.", "source": "https://www.preferred.jp/en/blog/tech/plamo-3-translate-weight-release", "source_label": "PFN release blog", "checked": "2026-11-10" }, { "kind": "checked", "label": "llama.cpp and mlx-lm support", "text": "llama.cpp master has a plamo3 architecture (src/models/plamo3.cpp), and the official GGUF header says general.architecture plamo3. With a sliding window set, llama.cpp loads a sliding-window pattern of 8: in each group of 8 layers, 7 use a 2,048-token sliding window and 1 uses full attention. For this size that is 3 full-attention and 21 sliding-window layers of 24, with 4 KV heads of size 128, which is what the context estimates use. mlx-lm master has mlx_lm/models/plamo3.py.", "source": "https://github.com/ggml-org/llama.cpp/blob/master/src/models/plamo3.cpp", "source_label": "llama.cpp plamo3.cpp", "checked": "2026-11-10" }, { "kind": "unknown", "label": "Not tested by us", "text": "We have not run any PLaMo 3 Translate model. No speed on this page is measured: tok/s figures are the site's bandwidth estimates. Translation quality is not tested by us, and the sizes are not compared with each other or with other models here.", "source": "https://huggingface.co/pfnet/plamo-3-translate-2b", "source_label": "Model card does not give this", "checked": "2026-11-10" } ], "kv": { "layers": 3, "kv_heads": 4, "head_dim": 128, "v_dim": 128, "swa_layers": 21, "swa_kv_heads": 4, "swa_head_dim": 128, "window": 2048, "arch": "swa", "max_ctx": 32768, "config_source": "https://huggingface.co/pfnet/plamo-3-translate-2b/blob/main/config.json", "config_note": "24 layers, window_size 2048, sliding_window_pattern 8: llama.cpp plamo3 (load_swa_pattern(ml, 8), LLAMA_SWA_TYPE_STANDARD) makes layer il sliding-window unless il % 8 == 7, so 3 full-attention + 21 sliding-window layers; the official GGUF metadata says the same (sliding_window 2048, sliding_window_pattern 8). max_ctx is the model card's 32,768-token training limit for input and output combined (the context field), not config.json's max_position_embeddings 262,144" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/pfnet/plamo-3-translate-2b", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" } ] }, { "id": "plamo-3-translate-8b", "name": "PLaMo 3 Translate 8B", "repo": "pfnet/plamo-3-translate-8b", "params_b": 8.53, "params_source": "https://huggingface.co/pfnet/plamo-3-translate-8b", "active_params_b": null, "active_source": null, "q4_file_gb": 5.28, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/pfnet/plamo-3-translate-8b-GGUF", "accessed": "2026-11-10", "license": "PLaMo Community License", "license_source": "https://plamo.preferredai.jp/info/plamo-community-license-en", "context": 32768, "released": "2026-11-05", "released_source": "https://www.preferred.jp/en/blog/tech/plamo-3-translate-weight-release", "task": "translation", "note": "Translation model, not a general chat assistant: it takes text in one language and returns a translation, through its own translation prompt format. Dense model, one of four PLaMo 3 Translate sizes. Context is 32,768 tokens because the model card says training used at most 32,768 tokens of input and output combined and asks you to split anything longer; config.json allows 262,144 positions, which this page does not treat as usable. Commercial use needs advance registration with Preferred Networks (see License below). Release date is the date of PFN's release blog post.", "quants": [ { "quant": "Q4_K_M", "bytes": 5280445728, "source": "https://huggingface.co/pfnet/plamo-3-translate-8b-GGUF", "by": "Preferred Networks" }, { "quant": "Q3_K_M", "bytes": 4310218016, "source": "https://huggingface.co/pfnet/plamo-3-translate-8b-GGUF", "by": "Preferred Networks" }, { "quant": "Q2_K", "bytes": 3450746144, "source": "https://huggingface.co/pfnet/plamo-3-translate-8b-GGUF", "by": "Preferred Networks" } ], "related": [ { "href": "/models/plamo-3-translate-0.8b/", "k": "Model", "t": "PLaMo 3 Translate 0.8B, another size of the same translation model" }, { "href": "/models/plamo-3-translate-2b/", "k": "Model", "t": "PLaMo 3 Translate 2B, another size of the same translation model" }, { "href": "/models/plamo-3-translate-31b/", "k": "Model", "t": "PLaMo 3 Translate 31B, another size of the same translation model" }, { "href": "/models/", "k": "Catalog", "t": "Every model in the catalog" } ], "checks": [ { "kind": "official", "label": "Model card", "text": "PFN's card describes PLaMo 3 Translate as a translation-specialized model, tagged translation on Hugging Face, not a general chat or coding assistant. It uses a translation chat template (the PFML3 prompt format). The release blog covers four sizes (0.8B, 2B, 8B, 31B) and 53 languages, and says the model accepts surrounding context and glossaries and can write plain or polite Japanese. The card says the longest training sequence was 32,768 tokens for input and output combined and to split anything longer, so this page uses 32,768 as the context, not config.json's 262,144. Safetensors total 8,531,750,912 parameters (8.53B), BF16, dense. The repo is not gated.", "source": "https://huggingface.co/pfnet/plamo-3-translate-8b", "source_label": "Model card", "checked": "2026-11-10" }, { "kind": "official", "label": "GGUF files", "text": "PFN publishes the GGUF build itself (pfnet/plamo-3-translate-8b-GGUF), one file per quant: Q4_K_M 5,280,445,728 bytes, Q3_K_M 4,310,218,016 bytes, Q2_K 3,450,746,144 bytes. This page's fit and speed estimates use the Q4_K_M file. The 31B card's example runs llama-server with -c 16384, --jinja and --no-prefill-assistant; 16,384 is only their example. We read the sizes from the Hugging Face file list and did not download them.", "source": "https://huggingface.co/pfnet/plamo-3-translate-8b-GGUF", "source_label": "GGUF repo", "checked": "2026-11-10" }, { "kind": "official", "label": "MLX 4-bit build", "text": "PFN also publishes an MLX 4-bit build for Apple silicon (pfnet/plamo-3-translate-8b-MLX-4bit), 6,066,104,664 bytes of safetensors. Its card says to run it with mlx-lm with trust_remote_code=True. Other official formats in the release are BF16, FP8, W4A16, W4AFP8 and NVFP4. We have not run the MLX build on a Mac, and no MLX speed is listed.", "source": "https://huggingface.co/pfnet/plamo-3-translate-8b-MLX-4bit", "source_label": "MLX-4bit repo", "checked": "2026-11-10" }, { "kind": "official", "label": "License: commercial use needs registration", "text": "PLaMo Community License. Personal and non-commercial use, including academic use, is free with no registration. Commercial use is free only if you (with your affiliates) had revenue of JPY 1 billion or less in the most recent fiscal year AND you register with PFN in advance through its form (https://forms.gle/mTL8tBLrMYXKNZD56). Above JPY 1 billion you need a separate license agreement with PFN (https://translate.preferredai.jp/llm/contact/). This is a summary of the license, not legal advice.", "source": "https://plamo.preferredai.jp/info/plamo-community-license-en", "source_label": "PLaMo Community License", "checked": "2026-11-10" }, { "kind": "vendor", "label": "PFN's benchmark claims", "text": "PFN's own claim: even the 0.8B model outperforms existing translation-specialized models (TranslateGemma, Hy-MT2) on Japanese-English and on Japanese to and from other languages, and PFN calls PLaMo 3 Translate the world's best translation-specialized LLM. Its benchmarks are PFTB and SaaS (PFN internal sets), WMT24++ (COMET-22) and FLORES+ (chrF++). PFN's card also says the Q2_K GGUF loses accuracy on some metrics while the other formats stay close to BF16. These are PFN's results; we have not reproduced any of them.", "source": "https://www.preferred.jp/en/blog/tech/plamo-3-translate-weight-release", "source_label": "PFN release blog", "checked": "2026-11-10" }, { "kind": "checked", "label": "llama.cpp and mlx-lm support", "text": "llama.cpp master has a plamo3 architecture (src/models/plamo3.cpp), and the official GGUF header says general.architecture plamo3. With a sliding window set, llama.cpp loads a sliding-window pattern of 8: in each group of 8 layers, 7 use a 2,048-token sliding window and 1 uses full attention. For this size that is 4 full-attention and 28 sliding-window layers of 32, with 4 KV heads of size 128, which is what the context estimates use. mlx-lm master has mlx_lm/models/plamo3.py.", "source": "https://github.com/ggml-org/llama.cpp/blob/master/src/models/plamo3.cpp", "source_label": "llama.cpp plamo3.cpp", "checked": "2026-11-10" }, { "kind": "unknown", "label": "Not tested by us", "text": "We have not run any PLaMo 3 Translate model. No speed on this page is measured: tok/s figures are the site's bandwidth estimates. Translation quality is not tested by us, and the sizes are not compared with each other or with other models here.", "source": "https://huggingface.co/pfnet/plamo-3-translate-8b", "source_label": "Model card does not give this", "checked": "2026-11-10" } ], "kv": { "layers": 4, "kv_heads": 4, "head_dim": 128, "v_dim": 128, "swa_layers": 28, "swa_kv_heads": 4, "swa_head_dim": 128, "window": 2048, "arch": "swa", "max_ctx": 32768, "config_source": "https://huggingface.co/pfnet/plamo-3-translate-8b/blob/main/config.json", "config_note": "32 layers, window_size 2048, sliding_window_pattern 8: llama.cpp plamo3 (load_swa_pattern(ml, 8), LLAMA_SWA_TYPE_STANDARD) makes layer il sliding-window unless il % 8 == 7, so 4 full-attention + 28 sliding-window layers; the official GGUF metadata says the same (sliding_window 2048, sliding_window_pattern 8). max_ctx is the model card's 32,768-token training limit for input and output combined (the context field), not config.json's max_position_embeddings 262,144" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/pfnet/plamo-3-translate-8b", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" } ] }, { "id": "plamo-3-translate-31b", "name": "PLaMo 3 Translate 31B", "repo": "pfnet/plamo-3-translate-31b", "params_b": 31.76, "params_source": "https://huggingface.co/pfnet/plamo-3-translate-31b", "active_params_b": null, "active_source": null, "q4_file_gb": 19.53, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/pfnet/plamo-3-translate-31b-GGUF", "accessed": "2026-11-10", "license": "PLaMo Community License", "license_source": "https://plamo.preferredai.jp/info/plamo-community-license-en", "context": 32768, "released": "2026-11-05", "released_source": "https://www.preferred.jp/en/blog/tech/plamo-3-translate-weight-release", "task": "translation", "note": "Translation model, not a general chat assistant: it takes text in one language and returns a translation, through its own translation prompt format. Dense model, one of four PLaMo 3 Translate sizes. Context is 32,768 tokens because the model card says training used at most 32,768 tokens of input and output combined and asks you to split anything longer; config.json allows 262,144 positions, which this page does not treat as usable. Commercial use needs advance registration with Preferred Networks (see License below). Release date is the date of PFN's release blog post.", "quants": [ { "quant": "Q4_K_M", "bytes": 19526593952, "source": "https://huggingface.co/pfnet/plamo-3-translate-31b-GGUF", "by": "Preferred Networks", "file": "plamo-3-translate-31b-Q4_K_M.gguf" }, { "quant": "Q3_K_M", "bytes": 15863131552, "source": "https://huggingface.co/pfnet/plamo-3-translate-31b-GGUF", "by": "Preferred Networks" }, { "quant": "Q2_K", "bytes": 12609924512, "source": "https://huggingface.co/pfnet/plamo-3-translate-31b-GGUF", "by": "Preferred Networks" } ], "related": [ { "href": "/models/plamo-3-translate-0.8b/", "k": "Model", "t": "PLaMo 3 Translate 0.8B, another size of the same translation model" }, { "href": "/models/plamo-3-translate-2b/", "k": "Model", "t": "PLaMo 3 Translate 2B, another size of the same translation model" }, { "href": "/models/plamo-3-translate-8b/", "k": "Model", "t": "PLaMo 3 Translate 8B, another size of the same translation model" }, { "href": "/models/", "k": "Catalog", "t": "Every model in the catalog" } ], "checks": [ { "kind": "official", "label": "Model card", "text": "PFN's card describes PLaMo 3 Translate as a translation-specialized model, tagged translation on Hugging Face, not a general chat or coding assistant. It uses a translation chat template (the PFML3 prompt format). The release blog covers four sizes (0.8B, 2B, 8B, 31B) and 53 languages, and says the model accepts surrounding context and glossaries and can write plain or polite Japanese. The card says the longest training sequence was 32,768 tokens for input and output combined and to split anything longer, so this page uses 32,768 as the context, not config.json's 262,144. Safetensors total 31,757,456,384 parameters (31.76B), BF16, dense. The repo is not gated.", "source": "https://huggingface.co/pfnet/plamo-3-translate-31b", "source_label": "Model card", "checked": "2026-11-10" }, { "kind": "official", "label": "GGUF files", "text": "PFN publishes the GGUF build itself (pfnet/plamo-3-translate-31b-GGUF), one file per quant: Q4_K_M 19,526,593,952 bytes, Q3_K_M 15,863,131,552 bytes, Q2_K 12,609,924,512 bytes. This page's fit and speed estimates use the Q4_K_M file. The 31B card's example runs llama-server with -c 16384, --jinja and --no-prefill-assistant; 16,384 is only their example. We read the sizes from the Hugging Face file list and did not download them.", "source": "https://huggingface.co/pfnet/plamo-3-translate-31b-GGUF", "source_label": "GGUF repo", "checked": "2026-11-10" }, { "kind": "official", "label": "MLX 4-bit build", "text": "PFN also publishes an MLX 4-bit build for Apple silicon (pfnet/plamo-3-translate-31b-MLX-4bit), 18,656,969,144 bytes of safetensors. Its card says to run it with mlx-lm with trust_remote_code=True. Other official formats in the release are BF16, FP8, W4A16, W4AFP8 and NVFP4. We have not run the MLX build on a Mac, and no MLX speed is listed.", "source": "https://huggingface.co/pfnet/plamo-3-translate-31b-MLX-4bit", "source_label": "MLX-4bit repo", "checked": "2026-11-10" }, { "kind": "official", "label": "License: commercial use needs registration", "text": "PLaMo Community License. Personal and non-commercial use, including academic use, is free with no registration. Commercial use is free only if you (with your affiliates) had revenue of JPY 1 billion or less in the most recent fiscal year AND you register with PFN in advance through its form (https://forms.gle/mTL8tBLrMYXKNZD56). Above JPY 1 billion you need a separate license agreement with PFN (https://translate.preferredai.jp/llm/contact/). This is a summary of the license, not legal advice.", "source": "https://plamo.preferredai.jp/info/plamo-community-license-en", "source_label": "PLaMo Community License", "checked": "2026-11-10" }, { "kind": "vendor", "label": "PFN's benchmark claims", "text": "PFN's own claim: even the 0.8B model outperforms existing translation-specialized models (TranslateGemma, Hy-MT2) on Japanese-English and on Japanese to and from other languages, and PFN calls PLaMo 3 Translate the world's best translation-specialized LLM. Its benchmarks are PFTB and SaaS (PFN internal sets), WMT24++ (COMET-22) and FLORES+ (chrF++). PFN's card also says the Q2_K GGUF loses accuracy on some metrics while the other formats stay close to BF16. These are PFN's results; we have not reproduced any of them.", "source": "https://www.preferred.jp/en/blog/tech/plamo-3-translate-weight-release", "source_label": "PFN release blog", "checked": "2026-11-10" }, { "kind": "checked", "label": "llama.cpp and mlx-lm support", "text": "llama.cpp master has a plamo3 architecture (src/models/plamo3.cpp), and the official GGUF header says general.architecture plamo3. With a sliding window set, llama.cpp loads a sliding-window pattern of 8: in each group of 8 layers, 7 use a 2,048-token sliding window and 1 uses full attention. For this size that is 8 full-attention and 56 sliding-window layers of 64, with 8 KV heads of size 128, which is what the context estimates use. mlx-lm master has mlx_lm/models/plamo3.py.", "source": "https://github.com/ggml-org/llama.cpp/blob/master/src/models/plamo3.cpp", "source_label": "llama.cpp plamo3.cpp", "checked": "2026-11-10" }, { "kind": "unknown", "label": "Not tested by us", "text": "We have not run any PLaMo 3 Translate model. No speed on this page is measured: tok/s figures are the site's bandwidth estimates. Translation quality is not tested by us, and the sizes are not compared with each other or with other models here.", "source": "https://huggingface.co/pfnet/plamo-3-translate-31b", "source_label": "Model card does not give this", "checked": "2026-11-10" } ], "kv": { "layers": 8, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "swa_layers": 56, "swa_kv_heads": 8, "swa_head_dim": 128, "window": 2048, "arch": "swa", "max_ctx": 32768, "config_source": "https://huggingface.co/pfnet/plamo-3-translate-31b/blob/main/config.json", "config_note": "64 layers, window_size 2048, sliding_window_pattern 8: llama.cpp plamo3 (load_swa_pattern(ml, 8), LLAMA_SWA_TYPE_STANDARD) makes layer il sliding-window unless il % 8 == 7, so 8 full-attention + 56 sliding-window layers; the official GGUF metadata says the same (sliding_window 2048, sliding_window_pattern 8). max_ctx is the model card's 32,768-token training limit for input and output combined (the context field), not config.json's max_position_embeddings 262,144" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/pfnet/plamo-3-translate-31b", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" } ] }, { "id": "twil-8b", "name": "webAI TwIL-8B", "repo": "webAI-Official/TwIL-8B", "params_b": 8.79, "params_source": "https://huggingface.co/webAI-Official/TwIL-8B", "active_params_b": null, "active_source": null, "q4_file_gb": 5.35, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/webAI-Official/TwIL-8B", "accessed": "2026-11-10", "license": "webAI Non-Commercial License ver. 1.0", "license_source": "https://huggingface.co/webAI-Official/TwIL-8B/blob/main/LICENSE.md", "context": 131072, "released": "2026-11-04", "released_source": "https://www.linkedin.com/posts/webai-official_activity-7523801223247101953-YG2S", "task": "math-code", "note": "Non-commercial use only: the webAI Non-Commercial License allows non-commercial research and educational purposes only. TwIL-8B is a reasoning fine-tune of IBM Granite 4.2 8B (same architecture and size) for contest math, code and instruction following, English only; its card says it is not a general assistant. The base Granite 4.2 8B stays Apache 2.0. Context is 131,072 tokens, inherited from the base. Release date is webAI's launch post; the Hugging Face repo's single commit is dated 2026-11-08.", "quants": [ { "quant": "Q4_K_M", "bytes": 5347917568, "source": "https://huggingface.co/webAI-Official/TwIL-8B", "by": "webAI", "file": "TwIL-8B-Q4_K_M-balanced.gguf" }, { "quant": "Q4_K_M", "bytes": 5347917248, "source": "https://huggingface.co/webAI-Official/TwIL-8B", "by": "webAI", "file": "TwIL-8B-Q4_K_M-verbose.gguf" }, { "quant": "Q8_0", "bytes": 9345613248, "source": "https://huggingface.co/webAI-Official/TwIL-8B", "by": "webAI", "file": "TwIL-8B-Q8_0-verbose.gguf" }, { "quant": "F16", "bytes": 17587420608, "source": "https://huggingface.co/webAI-Official/TwIL-8B", "by": "webAI", "file": "TwIL-8B-F16-verbose.gguf" } ], "related": [ { "href": "/models/granite-4.2-8b/", "k": "Base model", "t": "IBM Granite 4.2 8B, the Apache 2.0 base model TwIL-8B is trained from" }, { "href": "/models/", "k": "Catalog", "t": "Every model in the catalog" } ], "checks": [ { "kind": "official", "label": "Model card", "text": "webAI's card says TwIL-8B is post-trained from ibm-granite/granite-4.2-8b with full-parameter SFT and GRPO/DAPO reinforcement learning. It is a reasoning model that writes a block first, specialised in contest math, code and instruction following, English only, text in and text out. The card says it is not a general assistant, has no safety, preference or tool-use training beyond what Granite 4.2 carries, and is not a knowledge model. Safetensors total 8,791,592,960 parameters (8.79B), dense, GraniteForCausalLM: 40 layers, 32 attention heads, 8 KV heads of size 128, 131,072-token context. The repo is not gated.", "source": "https://huggingface.co/webAI-Official/TwIL-8B", "source_label": "Model card", "checked": "2026-11-10" }, { "kind": "official", "label": "Balanced and verbose variants", "text": "Two variants: balanced (the default, in the repo root) and verbose (in the verbose/ folder). GGUF files in the main repo: TwIL-8B-Q4_K_M-balanced.gguf 5,347,917,568 bytes (the card says 4.98 GiB, imatrix), TwIL-8B-Q4_K_M-verbose.gguf 5,347,917,248, TwIL-8B-Q8_0-verbose.gguf 9,345,613,248 and TwIL-8B-F16-verbose.gguf 17,587,420,608. The bf16 safetensors are 17,583,228,544 bytes per variant. This page's estimates use the balanced Q4_K_M file. The card's llama.cpp example is llama-cli -hf webAI-Official/TwIL-8B --hf-file TwIL-8B-Q4_K_M-balanced.gguf -cnv -c 40960 -n 32768, and it says answers need a large generation budget (webAI says balanced averages about 19.5k output tokens per AIME 2026 problem).", "source": "https://huggingface.co/webAI-Official/TwIL-8B/tree/main", "source_label": "Hugging Face file tree", "checked": "2026-11-10" }, { "kind": "official", "label": "License: non-commercial only", "text": "webAI Non-Commercial License ver. 1.0. Section 3.3: the work and any derivative works may be used or intended for use non-commercially only, and non-commercially means for non-commercial research and educational purposes only. So no commercial use. The base model, IBM Granite 4.2 8B, stays Apache 2.0. This is a summary of the license, not legal advice.", "source": "https://huggingface.co/webAI-Official/TwIL-8B/blob/main/LICENSE.md", "source_label": "LICENSE.md", "checked": "2026-11-10" }, { "kind": "vendor", "label": "webAI's benchmark claims", "text": "webAI's own figures (internal, high-effort pass@1, thinking on): MATH-500 98.2, AIME 2025 88.3, AIME 2026 88.3, GPQA Diamond 70.2, IFEval 91.8, IFBench 74.4 and LiveCodeBench 81.6 (the last two on the verbose variant; the card does not record the LiveCodeBench version). All were measured on bf16 weights, not on the GGUF files. The competitor figures in webAI's table are copied from other sources and were not re-run. These are webAI's claims; we have not reproduced any of them.", "source": "https://huggingface.co/webAI-Official/TwIL-8B", "source_label": "Model card", "checked": "2026-11-10" }, { "kind": "community", "label": "Community conversions", "text": "Community conversions exist and are not official, for example mradermacher/TwIL-8B-GGUF and pratikgx/TwIL-8B-MLX-4bit (4,945,853,960 bytes of safetensors). webAI publishes no MLX build. On a Mac, the official GGUF runs in llama.cpp, whose tagged releases support the Granite architecture; MLX only through a community conversion, which we have not run.", "source": "https://huggingface.co/models?search=TwIL-8B", "source_label": "Hugging Face search", "checked": "2026-11-10" }, { "kind": "unknown", "label": "Not tested by us", "text": "We have not run TwIL-8B. No speed on this page is measured: tok/s figures are the site's bandwidth estimates, and a reasoning model's long thinking output makes each answer take longer than the per-token speed suggests.", "source": "https://huggingface.co/webAI-Official/TwIL-8B", "source_label": "Model card does not give this", "checked": "2026-11-10" } ], "kv": { "layers": 40, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "arch": "gqa", "max_ctx": 131072, "config_source": "https://huggingface.co/webAI-Official/TwIL-8B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/webAI-Official/TwIL-8B", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" } ] }, { "id": "ornith-1.5-122b", "name": "Ornith-1.5-122B", "repo": "ornith-ai/Ornith-1.5-122B", "params_b": 125.09, "params_source": "https://huggingface.co/ornith-ai/Ornith-1.5-122B", "active_params_b": 10, "active_source": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B", "q4_file_gb": 78.92, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/Ornith-1.5-122B-GGUF", "accessed": "2026-11-11", "license": "MIT", "license_source": "https://huggingface.co/ornith-ai/Ornith-1.5-122B/blob/main/LICENSE", "context": 262144, "released": "2026-10-16", "released_source": "https://huggingface.co/ornith-ai/Ornith-1.5-122B", "note": "Ornith's card calls it a 122B MoE model and gives no active-parameter count. Its config.json has the same layout as Qwen3.5-122B-A10B, whose card says 10B activated, so this page uses 10B for its speed estimates. That is our inference from the matching config, not a figure Ornith publishes. Same architecture as Qwen3.5-122B-A10B (qwen35moe), so it runs in stock llama.cpp (bartowski quantized with release b11716). Fit ignores the 0.91 GB vision projector. Reasoning model, thinking on by default. On a Mac, Ornith's own MLX 4-bit build is listed under What we checked. Release date is the Hugging Face repo creation date.", "quants": [ { "quant": "Q4_K_M", "bytes": 75843911648, "source": "https://huggingface.co/ornith-ai/Ornith-1.5-122B-GGUF", "by": "ornith-ai (official)", "file": "Ornith-1.5-122B-Q4_K_M.gguf" }, { "quant": "Q4_K_M", "bytes": 78921736736, "source": "https://huggingface.co/bartowski/Ornith-1.5-122B-GGUF", "by": "bartowski (community)", "file": "Ornith-1.5-122B-Q4_K_M/ (2 parts)" }, { "quant": "Q5_K_M", "bytes": 88710135776, "source": "https://huggingface.co/ornith-ai/Ornith-1.5-122B-GGUF", "by": "ornith-ai (official)" }, { "quant": "Q6_K", "bytes": 102380498912, "source": "https://huggingface.co/ornith-ai/Ornith-1.5-122B-GGUF", "by": "ornith-ai (official)" }, { "quant": "Q8_0", "bytes": 132555710048, "source": "https://huggingface.co/ornith-ai/Ornith-1.5-122B-GGUF", "by": "ornith-ai (official)" }, { "quant": "BF16 mmproj", "bytes": 912263200, "source": "https://huggingface.co/ornith-ai/Ornith-1.5-122B-GGUF", "by": "ornith-ai (official)", "file": "mmproj-Ornith-1.5-122B-BF16.gguf", "projector": true } ], "related": [ { "href": "/models/qwen3.5-122b-a10b/", "k": "Model", "t": "Qwen3.5-122B-A10B: Ornith's card says Ornith builds on Qwen3.5, and the configs match" } ], "checks": [ { "kind": "official", "label": "Model card", "text": "Ornith's card calls Ornith-1.5-122B \"a 122B MOE model\" and does not state how many parameters are active per token. It says Ornith-1.0 was developed on top of Qwen3.5 and Gemma4, and Ornith-1.5 extends it. That is Ornith's description; it is not an official Qwen release, and the card does not say it is a fine-tune of the Qwen3.5-122B-A10B checkpoint. It is a reasoning model with thinking on by default. The card gives a context of up to 262,144 tokens. Hugging Face's safetensors metadata totals 125,086,497,008 BF16 parameters (125.09B on this page).", "source": "https://huggingface.co/ornith-ai/Ornith-1.5-122B", "source_label": "Model card", "checked": "2026-11-11" }, { "kind": "official", "label": "Weight files", "text": "The BF16 safetensors sum to 250,173,238,368 bytes (250.17 GB) at revision 7681ef608624e7532d8164128bcffdbdbaf1b1af. That is a download size, not GPU or unified memory.", "source": "https://huggingface.co/api/models/ornith-ai/Ornith-1.5-122B?blobs=true", "source_label": "Hugging Face API", "checked": "2026-11-11" }, { "kind": "official", "label": "Architecture config", "text": "config.json sets model_type qwen3_5_moe: 48 layers, with full attention on every 4th layer (12 attention layers) and linear attention on the rest; 2 KV heads with head dim 256; 256 experts with 8 routed per token, expert width 1,024; max_position_embeddings 262,144; a vision tower is included. Ornith's card gives no active-parameter count. This layout is the same as Qwen3.5-122B-A10B's, whose card says 10B activated, so this page uses 10B for its speed estimates. That is our inference from the matching config, not a figure Ornith publishes.", "source": "https://huggingface.co/ornith-ai/Ornith-1.5-122B/blob/main/config.json", "source_label": "config.json", "checked": "2026-11-11" }, { "kind": "checked", "label": "Same config as Qwen3.5-122B-A10B", "text": "We compared Ornith's config.json with Qwen/Qwen3.5-122B-A10B's. They are identical apart from cosmetic fields (dtype, transformers_version, and the vision model_type label), so the context and expert-split figures on this page use the same values as the Qwen3.5-122B-A10B page.", "source": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B/blob/main/config.json", "source_label": "Qwen config.json", "checked": "2026-11-11" }, { "kind": "official", "label": "MLX for Macs", "text": "Ornith publishes its own MLX builds. The 4-bit build is 69,612,579,840 bytes of safetensors (69.61 GB, or 64.83 GiB), including a 902,618,785-byte vision file. The 8-bit build (ornith-ai/Ornith-1.5-122B-MLX-8bit) is 130,648,354,386 bytes (130.65 GB). These are file sizes, not memory. The fit and speed figures on this page are for the GGUF, not these files.", "source": "https://huggingface.co/ornith-ai/Ornith-1.5-122B-MLX-4bit", "source_label": "MLX 4-bit repo", "checked": "2026-11-11" }, { "kind": "community", "label": "ezoushen MLX mixed 4.46 bpw", "text": "68,994,951,362 bytes of safetensors (68.99 GB). The uploader's card says it is 64.3 GiB with vision, 4.46 bits per weight, made for 128 GB Apple Silicon Macs (tested on an M5 Max), and runs with Splash (DFlash2 speculative decoding) and mlx-lm 0.31.3 (text only). Those are the uploader's claims; we have not run it.", "source": "https://huggingface.co/ezoushen/ornith-1.5-122b-mlx-mixed-4.46bpw", "source_label": "ezoushen card", "checked": "2026-11-11" }, { "kind": "official", "label": "Single-GPU serving claim", "text": "Ornith's own claim, not ours: the card says it \"serves on a single 80GB GPU\" with vLLM ≥ 0.19.1, SGLang ≥ 0.5.9 or Transformers ≥ 5.8.1. The BF16 weights are 250.17 GB, and Ornith's FP8 checkpoint is 132.95 GB (next row), so the 80 GB claim can only refer to a 4-bit-class checkpoint such as Ornith's NVFP4 build (77.63 GB). The card doesn't say which. We have not tested it.", "source": "https://huggingface.co/ornith-ai/Ornith-1.5-122B", "source_label": "Model card", "checked": "2026-11-11" }, { "kind": "official", "label": "FP8 checkpoint", "text": "ornith-ai/Ornith-1.5-122B-FP8 is 132,949,613,832 bytes (132.95 GB). A download size, not memory.", "source": "https://huggingface.co/ornith-ai/Ornith-1.5-122B-FP8", "source_label": "FP8 repo", "checked": "2026-11-11" }, { "kind": "official", "label": "NVFP4 checkpoint", "text": "ornith-ai/Ornith-1.5-122B-NVFP4 is 77,627,904,940 bytes (77.63 GB). A download size, not memory.", "source": "https://huggingface.co/ornith-ai/Ornith-1.5-122B-NVFP4", "source_label": "NVFP4 repo", "checked": "2026-11-11" }, { "kind": "community", "label": "llama.cpp", "text": "bartowski's GGUFs were made with llama.cpp release b11716 and use the qwen35moe architecture, so stock llama.cpp loads them. The Q4_K_M is two parts, 40,005,676,608 + 38,916,060,128 = 78,921,736,736 bytes (78.92 GB), the file the fit on this page uses. bartowski's BF16 vision projector is 912,263,520 bytes and is not counted in the fit.", "source": "https://huggingface.co/bartowski/Ornith-1.5-122B-GGUF", "source_label": "bartowski GGUF (commit a8bc08fa)", "checked": "2026-11-11" } ], "kv": { "layers": 12, "kv_heads": 2, "head_dim": 256, "v_dim": 256, "arch": "hybrid", "max_ctx": 262144, "config_source": "https://huggingface.co/ornith-ai/Ornith-1.5-122B/blob/main/config.json" }, "moe": { "expert_b": 115.964, "expert_active_b": 3.624, "layers": 48, "experts": 256, "experts_per_tok": 8, "expert_width": 1024, "config_source": "https://huggingface.co/ornith-ai/Ornith-1.5-122B/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/ornith-ai/Ornith-1.5-122B", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" }, { "source": "ollama", "url": "https://ollama.com/library/ornith-1.5:122b", "runs": "local", "pull": "ornith-1.5:122b" } ] }, { "id": "ripple-65b-preview", "name": "Ripple-65B Preview", "repo": "ricursiveAI/Ripple-65B-Preview", "params_b": 64.78, "params_source": "https://huggingface.co/ricursiveAI/Ripple-65B-Preview", "active_params_b": null, "active_source": null, "q4_file_gb": 38.8, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/AtomicChat/Ripple-65B-Preview-GGUF", "accessed": "2026-11-13", "license": "Apache 2.0", "license_source": "https://huggingface.co/ricursiveAI/Ripple-65B-Preview/blob/main/LICENSE", "context": 1048576, "released": "2026-11-04", "released_source": "https://huggingface.co/ricursiveAI/Ripple-65B-Preview", "note": "Ricursive Intelligence's dense 65B coding and agent model, trained from scratch, released as a preview. Dense means every token reads all of the weights, so per token it is slower than a mixture-of-experts model with a file of similar size. Ricursive calls the design \"thin-and-deep\" and \"Gemma 3N-like\": 80 layers, 64 of them sliding-window (1,024 tokens) and 16 full attention. Context is the card's 1,048,576 tokens. Q4_K_M is AtomicChat's plain llama.cpp quant (Ripple-65B-Preview-Q4_K_M.gguf, 38,798,820,576 bytes, no importance matrix); every size is in the table below. Stock llama.cpp loads it from build b12167, and tagged release v1.0.0 includes it. Release date is the Hugging Face repo creation date.", "quants": [ { "quant": "AD-IQ1_M", "bytes": 18493392352, "source": "https://huggingface.co/AtomicChat/Ripple-65B-Preview-GGUF", "by": "AtomicChat" }, { "quant": "AD-IQ2_XXS", "bytes": 20007273952, "source": "https://huggingface.co/AtomicChat/Ripple-65B-Preview-GGUF", "by": "AtomicChat" }, { "quant": "AD-IQ2_XS", "bytes": 22120482272, "source": "https://huggingface.co/AtomicChat/Ripple-65B-Preview-GGUF", "by": "AtomicChat" }, { "quant": "AD-IQ2_S", "bytes": 25310774752, "source": "https://huggingface.co/AtomicChat/Ripple-65B-Preview-GGUF", "by": "AtomicChat" }, { "quant": "AD-IQ3_XXS", "bytes": 27964982752, "source": "https://huggingface.co/AtomicChat/Ripple-65B-Preview-GGUF", "by": "AtomicChat" }, { "quant": "AD-IQ3_S", "bytes": 32411600352, "source": "https://huggingface.co/AtomicChat/Ripple-65B-Preview-GGUF", "by": "AtomicChat" }, { "quant": "AD-IQ4_XS", "bytes": 38722143712, "source": "https://huggingface.co/AtomicChat/Ripple-65B-Preview-GGUF", "by": "AtomicChat" }, { "quant": "Q4_K_M", "bytes": 38798820576, "source": "https://huggingface.co/AtomicChat/Ripple-65B-Preview-GGUF", "by": "AtomicChat", "file": "Ripple-65B-Preview-Q4_K_M.gguf" }, { "quant": "AD-Q4_K_M", "bytes": 40426079712, "source": "https://huggingface.co/AtomicChat/Ripple-65B-Preview-GGUF", "by": "AtomicChat" }, { "quant": "AD-Q5_K_M", "bytes": 47929624032, "source": "https://huggingface.co/AtomicChat/Ripple-65B-Preview-GGUF", "by": "AtomicChat" }, { "quant": "AD-Q6_K", "bytes": 57666225856, "source": "https://huggingface.co/AtomicChat/Ripple-65B-Preview-GGUF", "by": "AtomicChat", "file": "2 parts" }, { "quant": "Q8_0", "bytes": 68840113568, "source": "https://huggingface.co/AtomicChat/Ripple-65B-Preview-GGUF", "by": "AtomicChat", "file": "2 parts" }, { "quant": "BF16", "bytes": 129572324928, "source": "https://huggingface.co/AtomicChat/Ripple-65B-Preview-GGUF", "by": "AtomicChat", "file": "3 parts" } ], "weights": { "bytes": 129565519472, "shards": 16, "revision": "b172323ffd63e62de9ce1ba3830dfffe493c7671", "source": "https://huggingface.co/api/models/ricursiveAI/Ripple-65B-Preview?blobs=true", "note": "BF16 safetensors in 16 files. Commit b172323ffd63e62de9ce1ba3830dfffe493c7671." }, "checks": [ { "kind": "official", "label": "Model card", "text": "Ricursive's card calls it a \"65B parameter dense Transformer\", \"pretrained on 12 trillion tokens\", with a context of \"up to 1Mi (1,048,576 tokens)\", and describes the design as \"thin-and-deep\" and \"Gemma 3N-like\". config.json (model_type ricursive): 80 layers, hidden size 5,120, 80 attention heads, 8 KV heads, head size 128. Its layer_types list has 64 sliding_attention layers (window 1,024) and 16 full_attention layers (every 5th layer), with gated attention and value residuals. config.json's max_position_embeddings is 2,097,152; this page uses the card's 1,048,576. Hugging Face's safetensors total is 64,782,689,360 parameters (BF16). Transformers needs the repo's own code (configuration_ricursive.py and modeling_ricursive.py).", "source": "https://huggingface.co/ricursiveAI/Ripple-65B-Preview", "source_label": "Model card", "checked": "2026-11-13" }, { "kind": "official", "label": "License", "text": "Apache 2.0, in the repo's LICENSE file. This is a summary of the license, not legal advice.", "source": "https://huggingface.co/ricursiveAI/Ripple-65B-Preview/blob/main/LICENSE", "source_label": "LICENSE", "checked": "2026-11-13" }, { "kind": "official", "label": "Weight files", "text": "16 BF16 safetensors files sum to 129,565,519,472 bytes (129.57 GB) at commit b172323ffd63e62de9ce1ba3830dfffe493c7671. These are downloads, not the memory a runtime needs.", "source": "https://huggingface.co/api/models/ricursiveAI/Ripple-65B-Preview?blobs=true", "source_label": "Hugging Face API", "checked": "2026-11-13" }, { "kind": "official", "label": "NVFP4 checkpoint (experimental)", "text": "Ricursive's ricursiveAI/Ripple-65B-Preview-NVFP4, labeled \"(Experimental)\", is 57,213,852,024 bytes of safetensors (57.21 GB) at commit 2789217ba87d93e8d032fad25217c4c63a0e1d5b. Its card describes a \"conservative quantization ... NVFP4 for the MLP layers while keeping attention in BF16\". It is not a GGUF, and the fits on this page use the Q4_K_M file.", "source": "https://huggingface.co/ricursiveAI/Ripple-65B-Preview-NVFP4", "source_label": "NVFP4 repo", "checked": "2026-11-13" }, { "kind": "vendor", "label": "DGX Spark (Ricursive's claim)", "text": "Ricursive's card: \"At only 65B total parameters, the model can fit on consumer hardware such as the NVIDIA DGX Spark. To encourage local model development, we provide an NVFP4 checkpoint and colab\". That is Ricursive's claim. The DGX Spark rows on this page are our math on the Q4_K_M file, not a test.", "source": "https://huggingface.co/ricursiveAI/Ripple-65B-Preview", "source_label": "Model card", "checked": "2026-11-13" }, { "kind": "vendor", "label": "Benchmark scores (Ricursive's)", "text": "Ricursive's card reports Terminal-Bench 2.1 80.5 (marked self-reported) and SWE-bench Verified 76.5. The card itself says \"Replicating benchmark scores has been challenging\" and calls its table \"a best-effort attempt\". These are Ricursive's own reported scores, not checked by us.", "source": "https://huggingface.co/ricursiveAI/Ripple-65B-Preview", "source_label": "Model card", "checked": "2026-11-13" }, { "kind": "official", "label": "DFlash2 draft model", "text": "Ricursive publishes a DFlash2 draft model, ricursiveAI/Ripple-65B-Preview-DFlash2 (4,037,752,064 parameters, 8,075,513,104 bytes in BF16). It is only for speculative decoding (drafting) and is not a fit file. The speed estimates on this page do not assume drafting.", "source": "https://huggingface.co/ricursiveAI/Ripple-65B-Preview-DFlash2", "source_label": "DFlash2 repo", "checked": "2026-11-13" }, { "kind": "checked", "label": "DFlash2 in llama.cpp (not merged)", "text": "llama.cpp support for drafting Ripple with DFlash2 is not merged: pull request #31904 is open. A reviewer comment there says NoPE handling for the drafter was missing and that adding it improved acceptance on the reviewer's branch. The PR author's speed figures, with and without drafting, are under Reported by others below.", "source": "https://github.com/ggml-org/llama.cpp/pull/31904", "source_label": "llama.cpp PR #31904", "checked": "2026-11-13" }, { "kind": "community", "label": "GGUF files (AtomicChat)", "text": "AtomicChat/Ripple-65B-Preview-GGUF (commit df550a7178a87bbfb6ee2bdbd37603c553d536ee), GGUF architecture ricursive. AtomicChat's card calls the plain Q4_K_M a \"plain llama.cpp preset, no imatrix\". The AD files are AtomicChat's own per-tensor \"Atomic Dynamic\" layouts, made with an importance matrix. AtomicChat reports a mean KLD (how far the quant's output drifts from the original) of 0.0149 for AD-IQ4_XS against 0.0380 for the plain Q4_K_M; that is their measurement. This page uses the plain Q4_K_M. Every size is in the table below.", "source": "https://huggingface.co/AtomicChat/Ripple-65B-Preview-GGUF", "source_label": "AtomicChat GGUF card", "checked": "2026-11-13" }, { "kind": "community", "label": "GH200 run (AtomicChat's figure)", "text": "AtomicChat's card reports one run of the plain Q4_K_M with llama-cli on one GH200 96 GB: prompt 309 tokens/s, generation 40.6 tokens/s. That is AtomicChat's figure from a single run. We have not run it, and it is not a speed estimate on this page.", "source": "https://huggingface.co/AtomicChat/Ripple-65B-Preview-GGUF", "source_label": "AtomicChat GGUF card", "checked": "2026-11-13" }, { "kind": "checked", "label": "llama.cpp", "text": "Pull request #31692, \"model : add Ripple (ricursiveai)\", merged on 2026-11-05. The first build with it is b12167; AtomicChat's card says it \"Runs on stock llama.cpp b12167 or newer, no patch\". It is in tagged release v1.0.0 (2026-11-13). A loader is not a confirmed setup or a speed on this site.", "source": "https://github.com/ggml-org/llama.cpp/pull/31692", "source_label": "llama.cpp PR #31692", "checked": "2026-11-13" }, { "kind": "community", "label": "Ollama and LM Studio", "text": "AtomicChat's card says apps that bundle an older llama.cpp, LM Studio and Ollama among them, may not load it yet. We have not checked which llama.cpp build Ollama or LM Studio ships.", "source": "https://huggingface.co/AtomicChat/Ripple-65B-Preview-GGUF", "source_label": "AtomicChat GGUF card", "checked": "2026-11-13" }, { "kind": "community", "label": "MLX (unofficial)", "text": "There is no official or mlx-community MLX build. frogggias/Ripple-65B-Preview-MLX-4bit (commit 4a33b33f043f41a51ae7236edd440d64cfa7cc94, 36,442,941,052 bytes) is a community build; its card says \"mlx-lm doesn't know Ripple's architecture yet\", so it ships its own ricursive.py and needs --trust-remote-code. An 8-bit version exists (frogggias/Ripple-65B-Preview-MLX-8bit). MLX support for Ripple is unofficial and not merged into mlx-lm.", "source": "https://huggingface.co/frogggias/Ripple-65B-Preview-MLX-4bit", "source_label": "frogggias MLX card", "checked": "2026-11-13" } ], "related": [ { "href": "/gpus/nvidia-dgx-spark-128gb/", "k": "Hardware", "t": "NVIDIA DGX Spark (128GB), the machine Ricursive's card names" } ], "kv": { "layers": 16, "kv_heads": 8, "head_dim": 128, "v_dim": 128, "swa_layers": 64, "swa_kv_heads": 8, "swa_head_dim": 128, "window": 1024, "arch": "swa", "max_ctx": 1048576, "config_source": "https://huggingface.co/ricursiveAI/Ripple-65B-Preview/blob/main/config.json" }, "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/ricursiveAI/Ripple-65B-Preview", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" }, { "source": "ollama", "url": "https://ollama.com/library/ripple-preview:65b", "runs": "local", "pull": "ripple-preview:65b", "note": "Listed on Ollama as ripple-preview" } ] }, { "id": "inkling-2-small", "name": "Thinking Machines Inkling 2 Small", "repo": "thinkingmachines/Inkling-2-Small", "params_b": 593.72, "params_source": "https://huggingface.co/unsloth/Inkling-2-Small-GGUF", "active_params_b": 17, "active_source": "https://huggingface.co/thinkingmachines/Inkling-2-Small", "q4_file_gb": 354.15, "quant": "UD-Q4_K_XL", "q4_source": "https://huggingface.co/unsloth/Inkling-2-Small-GGUF", "accessed": "2026-11-13", "license": "Apache 2.0", "license_source": "https://huggingface.co/thinkingmachines/Inkling-2-Small", "context": 1048576, "released": "2026-11-12", "released_source": "https://thinkingmachines.ai/news/inkling-2-small/", "note": "Thinking Machines' second-generation Small model: a mixture-of-experts model with 594B total and 17B active parameters (the card's figures), text and image in, text out, trained in NVFP4. 593.72B is the total in unsloth's GGUF metadata (593,722,994,688); Hugging Face's own counter reads lower because NVFP4 packs two weights per byte. There is no plain Q4_K_M, so the size is unsloth's UD-Q4_K_XL (9 files, 354,150,089,945 bytes); every size is in the table below. Fit leaves out the 0.66 GB vision projector, which is extra if you send images. The checkpoint ships an MTP drafter for speculative decoding; the speed estimates on this page do not assume it. Release date is the day the weights went public.", "runtime_support": { "status": "llama.cpp pull request build and mlx-lm pull request", "text": "Stock llama.cpp, Ollama and LM Studio can't load Inkling 2 Small yet. The GGUFs use a new inkling2 architecture. Unsloth's guide (unsloth.ai/docs/models/inkling-2) builds llama.cpp from pull request #32187 (\"model : add Inkling 2 (stacked on Inkling 1)\", opened Nov 13, 2026, a draft and still open), and builds that support the original Inkling won't load it. On a Mac, mlx-community/Inkling-2-Small-4bit (uploaded Nov 13, 2026, 380,788,964,744 bytes) needs mlx-lm from unmerged pull request #2061 (\"Add Inkling 2 model\"). Its card says the routed experts keep NVFP4 (layers 0 to 39) or MXFP8 (layers 40 to 47) and the other weights are 4-bit, and it does not include the drafter. The fits and speeds on this page are math on the UD-Q4_K_XL file, not a test.", "source": "https://huggingface.co/unsloth/Inkling-2-Small-GGUF", "checked": "2026-11-13" }, "quants": [ { "quant": "UD-IQ1_S", "bytes": 171728965465, "source": "https://huggingface.co/unsloth/Inkling-2-Small-GGUF", "by": "unsloth", "file": "5 parts" }, { "quant": "UD-IQ1_M", "bytes": 178544971609, "source": "https://huggingface.co/unsloth/Inkling-2-Small-GGUF", "by": "unsloth", "file": "5 parts" }, { "quant": "UD-IQ2_XXS", "bytes": 190584459097, "source": "https://huggingface.co/unsloth/Inkling-2-Small-GGUF", "by": "unsloth", "file": "5 parts" }, { "quant": "UD-Q2_K_XL", "bytes": 204969271289, "source": "https://huggingface.co/unsloth/Inkling-2-Small-GGUF", "by": "unsloth", "file": "6 parts" }, { "quant": "UD-IQ3_XXS", "bytes": 225986223385, "source": "https://huggingface.co/unsloth/Inkling-2-Small-GGUF", "by": "unsloth", "file": "6 parts" }, { "quant": "UD-Q3_K_XL", "bytes": 285128675513, "source": "https://huggingface.co/unsloth/Inkling-2-Small-GGUF", "by": "unsloth", "file": "7 parts" }, { "quant": "UD-Q4_K_XL", "bytes": 354150089945, "source": "https://huggingface.co/unsloth/Inkling-2-Small-GGUF", "by": "unsloth", "file": "UD-Q4_K_XL/ (9 parts)" }, { "quant": "UD-Q8_K_XL", "bytes": 386510193913, "source": "https://huggingface.co/unsloth/Inkling-2-Small-GGUF", "by": "unsloth", "file": "9 parts" }, { "quant": "BF16", "bytes": 484382454681, "source": "https://huggingface.co/unsloth/Inkling-2-Small-GGUF", "by": "unsloth", "file": "11 parts" }, { "quant": "mmproj-F16", "bytes": 658223104, "source": "https://huggingface.co/unsloth/Inkling-2-Small-GGUF", "by": "unsloth", "file": "mmproj-F16.gguf", "projector": true } ], "weights": { "bytes": 393480084520, "shards": 25, "revision": "aac974122acd33ddb0d3e1f54433a9c5dd2249ce", "source": "https://huggingface.co/api/models/thinkingmachines/Inkling-2-Small?blobs=true", "note": "NVFP4 safetensors in 25 files, including mtp.safetensors (3,423,429,800 bytes), the MTP drafter. Commit aac974122acd33ddb0d3e1f54433a9c5dd2249ce." }, "checks": [ { "kind": "official", "label": "Model card", "text": "Thinking Machines' card: \"594B total, 17B active\", a context of \"up to 1M tokens\" (config model_max_length 1,048,576), text and image input with text output, and \"pretrained, midtrained and post-trained in NVFP4\". config.json (model_type inkling2): 48 layers, hidden size 4,608, 48 attention heads, 8 KV heads, head size 128; 40 local layers with a 1,024-token sliding window and 8 global layers; 384 routed experts, 8 per token, no shared expert; a vision tower; and an MTP drafter config with 5 layers. This page's parameter count is the GGUF metadata total of 593,722,994,688 in unsloth's GGUF. Hugging Face's safetensors counter shows 383.07B because NVFP4 packs two weights per byte; that is not the parameter count.", "source": "https://huggingface.co/thinkingmachines/Inkling-2-Small", "source_label": "Model card", "checked": "2026-11-13" }, { "kind": "official", "label": "Release and license", "text": "Thinking Machines' announcement is dated Nov 5, 2026 and carries the line \"Nov 12, 2026 Update: Inkling 2 Small's weights are now available for download on HuggingFace\". The release date on this page is that Nov 12 update. License: Apache 2.0, per the repo's license tag and the announcement. This is a summary of the license, not legal advice.", "source": "https://thinkingmachines.ai/news/inkling-2-small/", "source_label": "Thinking Machines announcement", "checked": "2026-11-13" }, { "kind": "official", "label": "Weight files", "text": "25 NVFP4 safetensors files sum to 393,480,084,520 bytes (393.48 GB) at commit aac974122acd33ddb0d3e1f54433a9c5dd2249ce, including mtp.safetensors (3,423,429,800 bytes), the MTP drafter. Thinking Machines also publishes an MXFP8 checkpoint, thinkingmachines/Inkling-2-Small-MXFP8; we did not read its size. These are downloads, not the memory a runtime needs.", "source": "https://huggingface.co/api/models/thinkingmachines/Inkling-2-Small?blobs=true", "source_label": "Hugging Face API", "checked": "2026-11-13" }, { "kind": "vendor", "label": "Server hardware (Thinking Machines' figures)", "text": "Thinking Machines' card says the NVFP4 checkpoint \"requires a GPU cluster with at least 500GB of aggregated VRAM\", for example 2x B300 or 4x H200, and that the MXFP8 checkpoint needs at least 800GB. For local inference it lists SGLang, vLLM, TokenSpeed and Unsloth. Those are Thinking Machines' statements about the safetensors checkpoints. The fits on this page use the UD-Q4_K_XL GGUF.", "source": "https://huggingface.co/thinkingmachines/Inkling-2-Small", "source_label": "Model card", "checked": "2026-11-13" }, { "kind": "vendor", "label": "Benchmark scores (Thinking Machines')", "text": "The card's eval results include SWE-bench Multilingual 81.2, GPQA Diamond 93 and MMMU Pro 81.6. These are Thinking Machines' own reported scores, not checked by us.", "source": "https://huggingface.co/thinkingmachines/Inkling-2-Small", "source_label": "Model card", "checked": "2026-11-13" }, { "kind": "community", "label": "GGUF files (unsloth)", "text": "unsloth/Inkling-2-Small-GGUF (commit caea5b5c473a199102fa9f9f32363e70faa49059). UD-Q4_K_XL is 9 files totaling 354,150,089,945 bytes, and the fits on this page use it. Unsloth's card says its BF16 GGUF keeps the original NVFP4 experts bit-exact (layers 0 to 39). The card's KLD table is Unsloth's measurement, and this page does not list it. mmproj-F16.gguf (658,223,104 bytes) is the vision projector. Every size is in the table below.", "source": "https://huggingface.co/unsloth/Inkling-2-Small-GGUF", "source_label": "unsloth GGUF card", "checked": "2026-11-13" }, { "kind": "checked", "label": "llama.cpp", "text": "Pull request #32187, \"model : add Inkling 2 (stacked on Inkling 1)\", was opened on 2026-11-13 as a draft and is still open. Unsloth's guide builds llama.cpp from it and says builds that support the original Inkling won't load Inkling 2. Stock llama.cpp, Ollama and LM Studio can't load it yet.", "source": "https://github.com/ggml-org/llama.cpp/pull/32187", "source_label": "llama.cpp PR #32187", "checked": "2026-11-13" }, { "kind": "community", "label": "MLX (unmerged mlx-lm pull request)", "text": "mlx-community/Inkling-2-Small-4bit was uploaded on 2026-11-13 (380,788,964,744 bytes). It needs mlx-lm from pull request #2061, \"Add Inkling 2 model\", which is open and not merged, so MLX support is unofficial for now. The card says the routed experts keep NVFP4 (layers 0 to 39) or MXFP8 (layers 40 to 47) and the other weights are 4-bit. It does not include the drafter.", "source": "https://github.com/ml-explore/mlx-lm/pull/2061", "source_label": "mlx-lm PR #2061", "checked": "2026-11-13" } ], "related": [ { "href": "/models/inkling-small/", "k": "Model", "t": "Inkling-Small, the previous model in Thinking Machines' Small line" } ], "kv_skip": "Inkling 2 Small's inkling2 architecture is only in an open llama.cpp pull request, so there is no verified llama.cpp KV formula for it yet. Its config.json marks the 40 sliding-window (1,024-token) layers in local_layer_ids, which the generic config reading here does not handle", "moe_skip": "config lacks expert width", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/thinkingmachines/Inkling-2-Small", "uploader": "other", "note": "Same account name as on Hugging Face; we found no link from the lab to it" }, { "source": "ollama", "url": "https://ollama.com/library/inkling-2-small:cloud", "runs": "cloud-only", "pull": "inkling-2-small:cloud" } ] }, { "id": "xing4.0-29b-a4b", "name": "Xing4.0-29B-A4B", "repo": "XingChen-AGI/Xing4.0-29B-A4B", "params_b": 31.22, "params_source": "https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B", "active_params_b": 4, "active_source": "https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B", "q4_file_gb": 20.06, "quant": "Q4_K_M", "q4_source": "https://huggingface.co/bartowski/XingChen-AGI_Xing4.0-29B-A4B-GGUF", "accessed": "2026-11-13", "license": "Apache 2.0", "license_source": "https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B/blob/main/LICENSE", "context": 262144, "released": "2026-09-16", "released_source": "https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B", "note": "China Telecom's agentic mixture-of-experts model, from its XingChen AGI team (the Xing series was formerly TeleChat): 29B total and 4B active per the card, with multi-head latent attention (MLA) and one MTP layer. The checkpoint holds 31.2B parameters; the card's 29B is its own rounding or count. It is new on this site because mainline llama.cpp added support on Nov 5, 2026, and tagged release v1.0.0 includes it. The official GGUF repo has no Q4_K_M, so the size is bartowski's Q4_K_M (XingChen-AGI_Xing4.0-29B-A4B-Q4_K_M.gguf, 20,056,603,584 bytes); every size is in the table below. Context is the card's 256K native window. Release date is the Hugging Face repo creation date.", "quants": [ { "quant": "iq2_M-slim", "bytes": 11800813760, "source": "https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B-GGUF", "by": "XingChen-AGI (official)" }, { "quant": "iq3_xxs", "bytes": 13989927648, "source": "https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B-GGUF", "by": "XingChen-AGI (official)" }, { "quant": "iq3_m", "bytes": 15446399712, "source": "https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B-GGUF", "by": "XingChen-AGI (official)" }, { "quant": "IQ4_XS", "bytes": 17982561216, "source": "https://huggingface.co/bartowski/XingChen-AGI_Xing4.0-29B-A4B-GGUF", "by": "bartowski (community)" }, { "quant": "iq4_xs", "bytes": 18272862944, "source": "https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B-GGUF", "by": "XingChen-AGI (official)" }, { "quant": "iq4_nl", "bytes": 19198485216, "source": "https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B-GGUF", "by": "XingChen-AGI (official)" }, { "quant": "Q4_K_M", "bytes": 20056603584, "source": "https://huggingface.co/bartowski/XingChen-AGI_Xing4.0-29B-A4B-GGUF", "by": "bartowski (community)", "file": "XingChen-AGI_Xing4.0-29B-A4B-Q4_K_M.gguf" }, { "quant": "IQ4_NL", "bytes": 20104012544, "source": "https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B-GGUF", "by": "XingChen-AGI (official)" }, { "quant": "q5_0", "bytes": 20416753312, "source": "https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B-GGUF", "by": "XingChen-AGI (official)" }, { "quant": "Q4_K_L", "bytes": 21686090688, "source": "https://huggingface.co/bartowski/XingChen-AGI_Xing4.0-29B-A4B-GGUF", "by": "bartowski (community)" }, { "quant": "q6_k", "bytes": 26876812000, "source": "https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B-GGUF", "by": "XingChen-AGI (official)" }, { "quant": "q8_0", "bytes": 34164155104, "source": "https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B-GGUF", "by": "XingChen-AGI (official)" } ], "weights": { "bytes": 62431093616, "shards": 12, "revision": "acbdebfe496d117e6fde3ce152d2450fb0d5fa53", "source": "https://huggingface.co/api/models/XingChen-AGI/Xing4.0-29B-A4B?blobs=true", "note": "BF16 safetensors in 12 files. Commit acbdebfe496d117e6fde3ce152d2450fb0d5fa53." }, "checks": [ { "kind": "official", "label": "Model card", "text": "The card says the model was \"developed by China Telecom Artificial Intelligence Technology Co., Ltd.\" (the Xing series was formerly TeleChat) and gives \"29B total / 4B active\" and a 256K native context, \"extensible to 512K\". config.json (model_type xing4_0): 40 layers with MLA (kv_lora_rank 512, qk_rope_head_dim 64), 64 routed experts with 4 per token plus 1 shared expert, the first 2 layers dense, 1 MTP (nextn) layer, and mHC (hc_mult 4). Hugging Face's safetensors total is 31,215,031,088 parameters (31,215,028,592 BF16 and 2,496 F32). The card's 29B is its own rounding or count; the checkpoint holds 31.2B. Transformers needs the repo's own model code.", "source": "https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B", "source_label": "Model card", "checked": "2026-11-13" }, { "kind": "official", "label": "License", "text": "Apache 2.0, in the repo's LICENSE file. This is a summary of the license, not legal advice.", "source": "https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B/blob/main/LICENSE", "source_label": "LICENSE", "checked": "2026-11-13" }, { "kind": "official", "label": "Weight files", "text": "12 BF16 safetensors files sum to 62,431,093,616 bytes (62.43 GB) at commit acbdebfe496d117e6fde3ce152d2450fb0d5fa53. These are downloads, not the memory a runtime needs.", "source": "https://huggingface.co/api/models/XingChen-AGI/Xing4.0-29B-A4B?blobs=true", "source_label": "Hugging Face API", "checked": "2026-11-13" }, { "kind": "official", "label": "Official GGUF files", "text": "XingChen-AGI/Xing4.0-29B-A4B-GGUF (commit 006904bc1d05ae2c37f938a4abdffee27e488544) has iq2_M-slim, iq3_xxs, iq3_m, iq4_xs, iq4_nl, q5_0, q6_k and q8_0, spelled as the repo spells them, and no Q4_K_M. It also has a second file named IQ4_NL (20,104,012,544 bytes) beside iq4_nl (19,198,485,216 bytes); both are in the table below.", "source": "https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B-GGUF", "source_label": "Official GGUF repo", "checked": "2026-11-13" }, { "kind": "community", "label": "bartowski GGUFs", "text": "bartowski/XingChen-AGI_Xing4.0-29B-A4B-GGUF (commit d324547035a19a4254afa224913a4287fb0dc0ee) was made with llama.cpp b12174, and the card says the MTP layers are kept for draft-mtp. The fits on this page use its Q4_K_M (XingChen-AGI_Xing4.0-29B-A4B-Q4_K_M.gguf, 20,056,603,584 bytes). Its IQ4_XS and Q4_K_L are in the table below.", "source": "https://huggingface.co/bartowski/XingChen-AGI_Xing4.0-29B-A4B-GGUF", "source_label": "bartowski GGUF card", "checked": "2026-11-13" }, { "kind": "checked", "label": "llama.cpp", "text": "Pull request #31589, \"Add support for the xing4_0\", merged on 2026-11-05 (build b12174 and later), and it is in tagged release v1.0.0 (2026-11-13). The maker's card, as of 2026-11-10, says \"Stock build b12174+ on CPU / CUDA\" and does not mention Metal or a Mac. We have not verified it on a Mac, and we have not checked Ollama or LM Studio. A loader is not a confirmed setup or a speed on this site.", "source": "https://github.com/ggml-org/llama.cpp/pull/31589", "source_label": "llama.cpp PR #31589", "checked": "2026-11-13" }, { "kind": "vendor", "label": "MTP drafting", "text": "The checkpoint has 1 MTP layer, and the GGUFs keep it, so llama.cpp can draft tokens with --spec-type draft-mtp. That is speculative decoding, not a speed figure, and the estimates on this page do not assume it.", "source": "https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B", "source_label": "Model card", "checked": "2026-11-13" }, { "kind": "vendor", "label": "SGLang and vLLM", "text": "SGLang mainline has supported it since 2026-11-05 (#39793), and vLLM runs it through an official out-of-tree plugin. Those are server setups; we have not run either.", "source": "https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B", "source_label": "Model card", "checked": "2026-11-13" }, { "kind": "vendor", "label": "Benchmark scores (XingChen's)", "text": "The card reports SWE-bench Verified 75.00 and Terminal-Bench 2.1 57.50. These are XingChen's own reported scores, not checked by us.", "source": "https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B", "source_label": "Model card", "checked": "2026-11-13" }, { "kind": "community", "label": "MLX (unofficial)", "text": "Stock mlx-lm has no xing4_0 class. mlx-community/Xing4.0-29B-A4B-OptiQ-4bit (20,579,605,322 bytes) says: \"Stock mlx-lm has no xing4_0 class — support is an unmerged pull request\". That repo works through the mlx-optiq package's vendored port, so MLX support is unofficial for now.", "source": "https://huggingface.co/mlx-community/Xing4.0-29B-A4B-OptiQ-4bit", "source_label": "mlx-community card", "checked": "2026-11-13" } ], "kv": { "layers": 40, "kv_heads": 1, "head_dim": 576, "v_dim": 0, "arch": "mla", "max_ctx": 262144, "config_source": "https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B/blob/main/config.json" }, "moe_skip": "derived active parameters (6.12B, 5.18B without embeddings) are more than 25% from the published 4B", "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/XingChen-AGI/Xing4.0-29B-A4B", "uploader": "official" } ] }, { "id": "edge0-145b", "name": "Edge0-145B", "repo": "Edge0/Edge0-145B-q3k", "params_b": 145.0, "params_source": "https://huggingface.co/Edge0/Edge0-145B-q3k", "active_params_b": 10.12, "active_source": "https://huggingface.co/Edge0/Edge0-145B-q3k", "accessed": "2026-11-14", "license": "Edge0 Community Model License 1.0 + Qwen Community License 1.0", "context": 262144, "released": "2026-11-12", "released_source": "https://edge0.ai/blog/introducing-edge0-145b-frontier-ai-capabilities-on-your-own-computer", "index": true, "weights": { "bytes": 105368210884, "files": "Q3_K build, all files in Edge0's own pack format", "source": "https://huggingface.co/api/models/Edge0/Edge0-145B-q3k?blobs=true", "note": "The NVFP4 build (Edge0/Edge0-145B-NVFP4) is 112,871,024,493 bytes (112.87 GB). In both builds the biggest files are experts.bin (40,076,574,720 bytes in Q3_K, 47,563,800,576 in NVFP4) and the same ple-fp8.gguf (51,200,246,112 bytes, the n-gram table Edge0's engine reads from the SSD). Neither build is a standard GGUF." }, "unknown_fit": { "lead": "An NVIDIA GPU with at least 16 GB, or an Apple Silicon Mac, running Edge0's own software.", "text": "Edge0's stated minimum is 16 GiB of GPU memory plus 20 GiB of system RAM. Edge0's main speed tests used a 24 GB card with about 31 to 38 GiB of RAM, and the files take about 105 to 113 GB of SSD. It won't run in llama.cpp, Ollama or LM Studio, or on AMD cards.", "short": "Edge0's stated minimum is 16 GiB of GPU memory plus 20 GiB of RAM." }, "note": "Edge0's mixture-of-experts model, derived from Qwen3.8-Flash-Next. 145B total includes a 51.2B n-gram embedding table (PLE, fp8) that Edge0's engine reads from the SSD, and a 1.26B MTP draft head. Each token uses 10.12B parameters (the card's figure; the repos' model.json says 6.6 B). Edge0 ships two builds: Q3_K, for NVIDIA CUDA 13 and Apple Silicon, and NVFP4, for CUDA only. Neither is a standard GGUF, so this site estimates no fit or speed. Every speed and score on this page is Edge0's own claim or a third party's report, never our estimate. Release date is the date on Edge0's blog post; the Hugging Face repos were created Nov 11, 2026 (UTC).", "kv_skip": "Edge0-145B ships as Edge0's own file pack with no standard Hugging Face config.json, and its layout (36 linear-attention layers and 12 indexed full-attention layers) has no verified llama.cpp KV formula. Edge0 says the KV cache is int8 with 32K resident", "runtime_support": { "tag": "Edge0 engine only", "status": "Edge0's own engine only (edge0-strata)", "text": "Stock llama.cpp, Ollama and LM Studio can't load Edge0-145B. The files are Edge0's own pack (experts.bin, labels.bin, latent_dense_*.bin, a model-skeleton.gguf that holds only attention, routing, shared experts and embeddings, and ple-fp8.gguf), not a standard GGUF. It runs on edge0-strata, Edge0's fork of Strata (MIT), started by edge0-infer serve or chat (Apache 2.0), or in the Edge0 Chat desktop app; edge0-infer's own native engine refuses these models. The engine needs an NVIDIA GPU with CUDA 13 or an Apple Silicon Mac with Metal (Q3_K build only). It has no AMD support: the fork removed Strata's AMD/HIP setup, so Radeon cards such as the RX 7800 XT can't run it. Ollama has no library entry for it, and we found no MLX conversion. This page estimates no fit or speed.", "source": "https://github.com/Edge0-AI/Edge0", "checked": "2026-11-14" }, "tables": [ { "id": "vendor-memory", "title": "Memory and disk, from Edge0's own figures", "text": "Edge0's residency figures, measured on its RTX 5090 D (24 GB) with residency on (\"disjoint expert sets, nothing counted twice\"). The 51.2 GB PLE table is read from the SSD on top of that (--ple-io direct by default). The whole model folder must be on a local disk. Edge0 gives no minimum SSD speed.", "columns": [ "Setup", "Build", "GPU memory", "System RAM", "Model folder on SSD" ], "rows": [ [ "RTX 5090 D (24 GB, a China-market card), residency on", "Q3_K", "23.14 GiB", "31.28 GiB", "about 105 GB" ], [ "RTX 5090 D (24 GB, a China-market card), residency on", "NVFP4", "23.25 GiB", "38.27 GiB", "about 113 GB" ], [ "Edge0's stated minimum, Linux and Windows", "Q3_K or NVFP4", "16 GiB", "20 GiB", "about 105 GB or 113 GB" ] ], "notes": [ "Low-RAM mode (--mmap-experts, or --resident-budget-gib N with --expert-profile) also reads uncached experts from the SSD. Edge0 has published a speed at that floor for the 145B Q3_K build only; no capped result exists for 145B NVFP4." ], "sources": [ { "label": "Model card", "url": "https://huggingface.co/Edge0/Edge0-145B-q3k" }, { "label": "Edge0 GitHub (edge0-infer/docs/edge0-145b.md)", "url": "https://github.com/Edge0-AI/Edge0" } ], "checked": "2026-11-14" }, { "id": "vendor-speeds", "title": "Speeds on Edge0's own table (Edge0's claim, with MTP speculative decoding)", "text": "Tokens per second, from Edge0's model card. These are Edge0's own numbers, not ours and not an estimate. Chat decode is three short questions (about 100 tokens out) and agent decode is the follow-up turns of a coding-agent-style session (at most 64 tokens out per turn), both with thinking off. Both decode columns use MTP speculative decoding: Edge0's shipped and bench configs turn it on (--spec 4 with the mtp/ draft layer). Edge0 publishes no plain-decoding number, and a third-party tester (The Henry Labs, running Edge0-145B) found that speculation can't be fully turned off on these packs.", "columns": [ "Machine (Edge0's label)", "Build", "4K prefill", "16K prefill", "Chat decode (MTP)", "Agent decode (MTP)" ], "rows": [ [ "RTX 5090 D (24 GB, a China-market card)", "Q3_K", "2169.9", "3554.9", "99.7", "121.2" ], [ "RTX 5090 D (24 GB, a China-market card)", "NVFP4", "1839.6", "3165.7", "79.6", "98.3" ], [ "RTX 4090 24 GB", "Q3_K", "2042.0", "3179.4", "69.9", "91.7" ], [ "RTX Spark N1X 60 GiB", "Q3_K", "674.0", "770.3", "37.7", "36.7" ], [ "RTX Spark N1X 60 GiB", "NVFP4", "821.2", "922.4", "40.1", "46.0" ], [ "Apple M2 Ultra (Metal)", "Q3_K", "755.6", "800.3", "60.9", "57.0" ], [ "The same RTX 5090 D, capped to 16 GiB VRAM + 20 GiB RAM", "Q3_K", "249", "335", "36.8", "44.9" ] ], "notes": [ "RTX 5090 D (24 GB) is the China-market GeForce RTX 5090 D v2, not the 32 GB RTX 5090 sold in the US. US readers can compare with the RTX 4090 row.", "\"RTX Spark N1X 60 GiB\" is Edge0's label. We have no hardware entry that matches it.", "Capped rows: Edge0 capped VRAM with --vram-reserve-mib 8192 and system RAM with a cgroup MemoryMax=20G, no swap. Edge0 says: \"These are test limits, not minimum system requirements.\"" ], "sources": [ { "label": "Model card", "url": "https://huggingface.co/Edge0/Edge0-145B-q3k" }, { "label": "Edge0 GitHub (docs/models/edge0-145b.md)", "url": "https://github.com/Edge0-AI/Edge0" } ], "checked": "2026-11-14" }, { "id": "claims-vs-runs", "title": "Edge0's headline vs its own card vs two independent runs", "text": "The 29 GB headline doesn't match the card's VRAM + RAM figures. Plan from the card's figures. The card's and the two runs' decode figures all have speculative decoding on, and Edge0's blog ties its 80+ figure to speculative decoding; none is a plain-decoding speed. The setups, builds and tests differ, so the speeds don't line up with each other.", "columns": [ "Source", "Setup", "Memory", "Decode" ], "rows": [ [ "Edge0's blog and launch post (marketing)", "Hardware not named", "\"29GB peak memory\"", "\"80+ tokens/s\" (blog); \"75+ toks/s\" (X post)" ], [ "Edge0's model card", "RTX 5090 D (24 GB, a China-market card), Q3_K", "23.14 GiB VRAM + 31.28 GiB RAM (about 54 GiB)", "99.7 chat / 121.2 agent tok/s (MTP)" ], [ "Edge0's model card", "RTX 5090 D (24 GB, a China-market card), NVFP4", "23.25 GiB VRAM + 38.27 GiB RAM (about 62 GiB)", "79.6 chat / 98.3 agent tok/s (MTP)" ], [ "The Henry Labs (third party)", "RTX 3090 24 GB, Q3_K, cloud host with 116.4 GiB RAM", "Not reported", "54.1 chat / 57.3 agent tok/s with Edge0's bench tool (MTP, --spec 4)" ], [ "Dylan Sun (third party)", "RTX 5090 32 GB (not the 5090 D), NVFP4, Windows 11, 128 GB RAM", "GPU peak 29,343 MiB; engine RSS peak 38,492 MiB", "104.7 tok/s code generation, 1,024-token outputs (MTP, window 4)" ] ], "notes": [ "The two third-party runs are listed in full under Reported by others below. Self-reported, not independently reproduced." ], "sources": [ { "label": "Edge0 blog", "url": "https://edge0.ai/blog/introducing-edge0-145b-frontier-ai-capabilities-on-your-own-computer" }, { "label": "Model card", "url": "https://huggingface.co/Edge0/Edge0-145B-q3k" }, { "label": "The Henry Labs REPORT.md", "url": "https://github.com/thehenrylabsdev/edge0-145b-rtx3090-benchmark/blob/main/REPORT.md" }, { "label": "Dylan Sun report", "url": "https://github.com/Dylan997S/benchmarks-strata/tree/main/reports/2026-11-14-rtx5090-edge0-vs-strata-iq4" } ], "checked": "2026-11-14" }, { "id": "vendor-benchmarks", "title": "Benchmark scores (Edge0's internal evaluations)", "text": "Edge0's own table, from its model card and GitHub docs. Edge0 says the Edge0, Qwen, DeepSeek and Claude results all come from its internal evaluations, and that the setup and reproduction steps will come with a technical report that is not out yet. We have not checked these scores. On this table the base model Qwen3.8-Flash-Next scores above Edge0-145B on GPQA-Diamond, IFBench and HLE; Edge0-145B is above it on DeepSWE, Terminal-Bench and AA-LCR.", "columns": [ "Model", "DeepSWE 1.1", "Terminal-Bench 2.1", "AA-LCR", "IFBench", "GPQA-Diamond", "HLE" ], "rows": [ [ "Edge0-145B", "64.6", "86.5", "79.3", "80.6", "90.4", "35.4" ], [ "Edge0-120B", "51.3", "75.3", "81.0", "79.3", "87.4", "31.0" ], [ "DeepSeek-V4-Flash-0731", "54.4", "82.7", "79.7", "79.2", "90.8", "33.8" ], [ "Qwen3.8-Flash-Next", "58.7", "83.3", "79.1", "81.3", "91.7", "35.9" ], [ "Qwen3.8-27B", "42.2", "73.0", "82.0", "79.5", "89.2", "30.8" ], [ "Claude-Opus-4.6 (Max)", "—", "78.2", "78.0", "62.5", "91.3", "40.0" ] ], "notes": [ "Edge0's blog gives Terminal-Bench 86.52 for Edge0-145B." ], "sources": [ { "label": "Model card", "url": "https://huggingface.co/Edge0/Edge0-145B-q3k" }, { "label": "Edge0 blog", "url": "https://edge0.ai/blog/introducing-edge0-145b-frontier-ai-capabilities-on-your-own-computer" } ], "checked": "2026-11-14" } ], "checks": [ { "kind": "official", "label": "Model card", "text": "Edge0's card: 145B total parameters and 10.12B active per token. 48 layers: 36 linear-attention layers, with an indexed full-attention layer every 4th. 512 routed experts per layer plus 1 shared, 10 routed experts active per token (SwiGLU width 640). The expert pool is 84.6B parameters at a latent width of 1792. A 51.2B n-gram embedding table (PLE) is stored in fp8 and memory-mapped from the SSD, and a 1.26B MTP head is the draft model for speculative decoding. The 51.2B table counts inside the 145B total. Context 256K, with an int8 KV cache and 32K resident. Edge0 gives the expert-weight builds as Q3_K 40.81 GiB and NVFP4 47.79 GiB; those are not the whole download. Edge0 describes the model as \"trained natively at 4 bits\" and derived from Qwen3.8-Flash-Next; its release-scope.json says \"No training-source commit or independent ownership audit is asserted\".", "source": "https://huggingface.co/Edge0/Edge0-145B-q3k", "source_label": "Model card", "checked": "2026-11-14" }, { "kind": "official", "label": "Weight files", "text": "Hugging Face API totals: Edge0/Edge0-145B-q3k is 105,368,210,884 bytes (105.37 GB) and Edge0/Edge0-145B-NVFP4 is 112,871,024,493 bytes (112.87 GB). NVFP4 files: experts.bin 47,563,800,576; ple-fp8.gguf 51,200,246,112; model-skeleton.gguf 5,991,569,248; latent_dense_q8.bin 3,743,416,320; latent_dense_q4.bin 1,981,808,640; dense.bin 1,538,625,024; mtp/experts.bin 707,788,800; mtp/dense.bin 116,099,072 bytes. Q3_K has a smaller experts.bin (40,076,574,720 bytes), the same PLE, skeleton, dense and latent files, and a scales/ folder of 48 small files. Edge0's \"about 113GB\" for the NVFP4 repo matches. These are downloads, not the memory the engine needs.", "source": "https://huggingface.co/api/models/Edge0/Edge0-145B-q3k?blobs=true", "source_label": "Hugging Face API", "checked": "2026-11-14" }, { "kind": "official", "label": "Requirements (Edge0's docs)", "text": "Edge0's docs: \"Linux or Windows with an NVIDIA GPU and CUDA 13, or macOS on Apple Silicon with Metal. The macOS measurements were taken on an Apple M2 Ultra.\" and \"On Linux and Windows, at least 16 GiB of GPU memory and 20 GiB of system memory.\" The same docs also say the constrained-memory speed rows are measured test budgets, not general minimum requirements. Building edge0-strata needs the CUDA 13 toolkit (nvcc), and Visual Studio 2022 on Windows x64 for CUDA 13.0 to 13.2; Strata's install doc says CUDA 13 packages need NVIDIA driver 580 or newer. On a Mac only the Q3_K folder runs (NVFP4 is CUDA only), the Metal template sets a 32,768-token context, and the first start writes mtp-q2_0.gguf into the model folder, so the folder must be writable. The Edge0 Chat 0.2.0 app (Nov 11, 2026) has macos-arm64, windows-x64 and windows-arm64 builds.", "source": "https://github.com/Edge0-AI/Edge0", "source_label": "Edge0 GitHub (edge0-infer/docs/edge0-145b.md)", "checked": "2026-11-14" }, { "kind": "official", "label": "AMD", "text": "Edge0's engine has no AMD support. The edge0-strata fork removed upstream Strata's AMD/HIP, Intel SYCL and Docker setups (its EDGE0.md: \"Removed because neither the Linux CUDA build nor the server uses them\"), and the build instructions cover only NVIDIA CUDA 13 and Apple Metal. So a Radeon card such as the RX 7800 XT can't run it with Edge0's engine.", "source": "https://github.com/Edge0-AI/Edge0", "source_label": "Edge0 GitHub (edge0-strata/EDGE0.md)", "checked": "2026-11-14" }, { "kind": "official", "label": "Speculative decoding in Edge0's configs", "text": "Every shipped config uses MTP speculative decoding. The model.json in each repo and the CUDA template (edge0-145b.json) set --spec 4 --spec-min-p 0.5 --mtp ./mtp, the Metal templates also set --spec 4, and edge0-infer's docs list mtp/ (\"Draft layer for speculative decoding\") as a required file. The bench config (model-bench.json in the 145B Q3_K repo) also sets --spec 4 with --mtp. Edge0's blog ties its \"80+ tokens/s\" figure to speculative decoding. Edge0 publishes no plain-decoding number, and this site does not estimate one.", "source": "https://huggingface.co/Edge0/Edge0-145B-q3k", "source_label": "Hugging Face repo (model.json)", "checked": "2026-11-14" }, { "kind": "checked", "label": "Which RTX 5090 D", "text": "Edge0's table names an \"RTX 5090 D 24 GB\". NVIDIA's China site lists the GeForce RTX 5090 D v2 with 24 GB of GDDR7 on a 384-bit bus. It is a China-market card, not sold in the US, and not the 32 GB GeForce RTX 5090 on this site, so Edge0's numbers are not attached to that page. US readers can compare with Edge0's RTX 4090 rows.", "source": "https://www.nvidia.cn/geforce/graphics-cards/50-series/rtx-5090-d-v2/", "source_label": "NVIDIA China, RTX 5090 D v2", "checked": "2026-11-14" }, { "kind": "vendor", "label": "Edge0's headline claims", "text": "Edge0's blog says \"29GB peak memory\", that the model \"runs on a compatible personal laptop, such as a MacBook Pro\", \"80+ tokens/s decode\", \"1,100+ tokens/s prefill\", \"approximately 120W typical power\" and \"enabling local use on a 48GB laptop\". Its launch post on X said it \"needs just 29G memory\" and \"decodes 75+ toks/s, prefills 1050+ toks/s on RTX Spark and Mac\". The blog doesn't name the hardware behind the 29 GB figure; it says that will come with a technical report. That headline doesn't match the card: Edge0's own residency figures on a 24 GB card are 23.14 GiB VRAM + 31.28 GiB RAM (Q3_K, about 54 GiB together) and 23.25 GiB + 38.27 GiB (NVFP4, about 62 GiB). Plan from the card's VRAM and RAM figures, not the 29 GB headline. These are Edge0's marketing claims; we have not reproduced them.", "source": "https://edge0.ai/blog/introducing-edge0-145b-frontier-ai-capabilities-on-your-own-computer", "source_label": "Edge0 blog", "checked": "2026-11-14" }, { "kind": "official", "label": "Where Edge0's own figures disagree", "text": "Each repo's model.json says \"active_params\": \"6.6 B\", while the card and GitHub docs say 10.12B; this page uses the card's 10.12B. The card says 256K context, but the shipped CUDA template (edge0-145b.json) sets --max-context 131072 (128K) and the macOS Metal template sets 32768; Edge0's blog says \"128K–256K context\". The context shown on this page is the card's 262,144; plan on 131,072 with the shipped CUDA template.", "source": "https://github.com/Edge0-AI/Edge0", "source_label": "Edge0 GitHub and model.json", "checked": "2026-11-14" }, { "kind": "official", "label": "License", "text": "Two licenses apply to the weights: the Edge0 Community Model License 1.0 (dated 12 Nov 2026) and, on top, the Qwen Community License 1.0 of the base model (shipped as third-party/Qwen-LICENSE.txt). Free under Edge0's license: personal use and local deployment; research, education, benchmarking and publishing results; modifying, fine-tuning, quantizing, converting and non-commercial sharing; internal use inside a business; paid work by individuals that doesn't give customers access to the model; and commercial use of the outputs. You need a separate written license from Edge0 (info@edge0.ai) to sell or rent the model, weights, adapters or paid bundles built around it; to run a paid model service (an API or hosted endpoint) for third parties; or to sell an independent AI coding or office-assistant product built on it, even one that runs on the user's device. The Qwen license adds its own terms: a Model-as-a-Service or AI Work Assistant business needs a separate commercial license from Qwen, and a product with more than 100 million monthly active users or more than US$20 million in monthly revenue must show the model name. Edge0's permission does not replace Qwen's. Edge0's code (edge0-infer and the rest of the framework) is Apache 2.0, and edge0-strata keeps Strata's MIT license. This is a summary of the licenses, not legal advice.", "source": "https://huggingface.co/Edge0/Edge0-145B-q3k", "source_label": "Model card", "checked": "2026-11-14" }, { "kind": "community", "label": "Stock llama.cpp", "text": "The Henry Labs' RTX 3090 report says: \"Stock llama.cpp cannot load this model: the pack keeps its experts outside the GGUF in a latent layout that upstream's qwen4exp loader does not support.\"", "source": "https://github.com/thehenrylabsdev/edge0-145b-rtx3090-benchmark/blob/main/REPORT.md", "source_label": "The Henry Labs REPORT.md", "checked": "2026-11-14" }, { "kind": "checked", "label": "Ollama and MLX", "text": "Ollama's library pages /library/edge0, /library/edge0-145b and /library/edge0-120b return 404. We found no MLX conversion.", "source": "https://ollama.com/library/edge0", "source_label": "Ollama library", "checked": "2026-11-14" }, { "kind": "unknown", "label": "Not documented", "text": "Unknown, and not estimated here: a plain-decoding (no speculation) speed, a minimum SSD speed, the hardware behind Edge0's 29 GB headline, Edge0's evaluation setup (promised with a technical report that is not out yet), speeds for long outputs, and any fit or speed on a PC or Mac in our catalog.", "source": "https://edge0.ai/blog/introducing-edge0-145b-frontier-ai-capabilities-on-your-own-computer", "source_label": "Not in Edge0's blog or card", "checked": "2026-11-14" } ], "related": [ { "href": "/models/edge0-120b/", "k": "Model", "t": "Edge0-120B, the smaller Edge0 model on the same engine" }, { "href": "/models/qwen3.8-flash-next/", "k": "Model", "t": "Qwen3.8-Flash-Next, the base model Edge0 derived it from" }, { "href": "/guides/dgx-spark-cluster-two-sparks/#ssd-streamed-setups-on-one-spark", "k": "Guide", "t": "How Strata-style SSD streaming works and what we've seen on a Spark (Edge0's engine is a Strata fork)" }, { "href": "/guides/hugging-face-alternatives-download-models/", "k": "Guide", "t": "Other places to download models besides Hugging Face" } ], "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Edge0/Edge0-145B-q3k", "uploader": "official", "note": "Q3_K build" }, { "source": "modelscope", "url": "https://modelscope.cn/models/Edge0/Edge0-145B-NVFP4", "uploader": "official", "note": "NVFP4 build" } ] }, { "id": "edge0-120b", "name": "Edge0-120B", "repo": "Edge0/Edge0-120B-q3k", "params_b": 120.0, "params_source": "https://huggingface.co/Edge0/Edge0-120B-q3k", "active_params_b": 8.64, "active_source": "https://huggingface.co/Edge0/Edge0-120B-q3k", "accessed": "2026-11-14", "license": "Edge0 Community Model License 1.0 + Qwen Community License 1.0", "context": 262144, "released": "2026-11-12", "released_source": "https://edge0.ai/blog/introducing-edge0-145b-frontier-ai-capabilities-on-your-own-computer", "index": true, "weights": { "bytes": 92281883893, "files": "Q3_K build, all files in Edge0's own pack format", "source": "https://huggingface.co/api/models/Edge0/Edge0-120B-q3k?blobs=true", "note": "ModelScope's size for the same repo. The NVFP4 build (Edge0/Edge0-120B-NVFP4) is 97,629,769,542 bytes (97.63 GB). In both builds the biggest files are experts.bin (28,626,124,800 bytes in Q3_K, 33,974,255,616 in NVFP4) and the same ple-fp8.gguf (51,200,246,112 bytes, the n-gram table Edge0's engine reads from the SSD). Neither build is a standard GGUF." }, "unknown_fit": { "lead": "An NVIDIA GPU with at least 16 GB, or an Apple Silicon Mac, running Edge0's own software.", "text": "Edge0's stated minimum is 16 GiB of GPU memory plus 20 GiB of system RAM. Edge0's main speed tests used a 24 GB card with about 18 to 23 GiB of RAM, and the files take about 92 to 98 GB of SSD. It won't run in llama.cpp, Ollama or LM Studio, or on AMD cards.", "short": "Edge0's stated minimum is 16 GiB of GPU memory plus 20 GiB of RAM." }, "note": "Edge0's mixture-of-experts model, derived from Qwen3.8-Flash-Next. 120B total includes a 51.2B n-gram embedding table (PLE, fp8) that Edge0's engine reads from the SSD, and a 1.26B MTP draft head. Each token uses 8.64B parameters (the card's figure; the repos' model.json says 6.1 B). Edge0 ships two builds: Q3_K, for NVIDIA CUDA 13 and Apple Silicon, and NVFP4, for CUDA only. Neither is a standard GGUF, so this site estimates no fit or speed. Every speed and score on this page is Edge0's own claim or a third party's report, never our estimate. Release date is the date on Edge0's blog post; the Hugging Face repos were created Nov 11, 2026 (UTC).", "kv_skip": "Edge0-120B ships as Edge0's own file pack with no standard Hugging Face config.json, and its layout (36 linear-attention layers and 12 indexed full-attention layers) has no verified llama.cpp KV formula", "runtime_support": { "tag": "Edge0 engine only", "status": "Edge0's own engine only (edge0-strata)", "text": "Stock llama.cpp, Ollama and LM Studio can't load Edge0-120B. The files are Edge0's own pack (experts.bin, labels.bin, latent_dense_*.bin, a model-skeleton.gguf that holds only attention, routing, shared experts and embeddings, and ple-fp8.gguf), not a standard GGUF. It runs on edge0-strata, Edge0's fork of Strata (MIT), started by edge0-infer serve or chat (Apache 2.0), or in the Edge0 Chat desktop app; edge0-infer's own native engine refuses these models. The engine needs an NVIDIA GPU with CUDA 13 or an Apple Silicon Mac with Metal (Q3_K build only). It has no AMD support: the fork removed Strata's AMD/HIP setup, so Radeon cards such as the RX 7800 XT can't run it. Ollama has no library entry for it, and we found no MLX conversion. This page estimates no fit or speed.", "source": "https://github.com/Edge0-AI/Edge0", "checked": "2026-11-14" }, "tables": [ { "id": "vendor-memory", "title": "Memory and disk, from Edge0's own figures", "text": "Edge0's residency figures, measured on its RTX 5090 D (24 GB) with residency on (\"disjoint expert sets, nothing counted twice\"). The 51.2 GB PLE table is read from the SSD on top of that (--ple-io direct by default). The whole model folder must be on a local disk. Edge0 gives no minimum SSD speed.", "columns": [ "Setup", "Build", "GPU memory", "System RAM", "Model folder on SSD" ], "rows": [ [ "RTX 5090 D (24 GB, a China-market card), residency on", "Q3_K", "22.93 GiB", "18.40 GiB", "about 92 GB" ], [ "RTX 5090 D (24 GB, a China-market card), residency on", "NVFP4", "23.02 GiB", "23.40 GiB", "about 98 GB" ], [ "Edge0's stated minimum, Linux and Windows", "Q3_K or NVFP4", "16 GiB", "20 GiB", "about 92 GB or 98 GB" ] ], "notes": [ "Low-RAM mode (--mmap-experts, or --resident-budget-gib N with --expert-profile) also reads uncached experts from the SSD. Edge0 has published speeds at that floor for both 120B builds." ], "sources": [ { "label": "Model card", "url": "https://huggingface.co/Edge0/Edge0-120B-q3k" }, { "label": "Edge0 GitHub (edge0-infer/docs/edge0-120b.md)", "url": "https://github.com/Edge0-AI/Edge0" } ], "checked": "2026-11-14" }, { "id": "vendor-speeds", "title": "Speeds on Edge0's own table (Edge0's claim, with MTP speculative decoding)", "text": "Tokens per second, from Edge0's model card. These are Edge0's own numbers, not ours and not an estimate. Chat decode is three short questions (about 100 tokens out) and agent decode is the follow-up turns of a coding-agent-style session (at most 64 tokens out per turn), both with thinking off. Both decode columns use MTP speculative decoding: Edge0's shipped and bench configs turn it on (--spec 4 with the mtp/ draft layer). Edge0 publishes no plain-decoding number, and a third-party tester (The Henry Labs, running Edge0-145B) found that speculation can't be fully turned off on these packs.", "columns": [ "Machine (Edge0's label)", "Build", "4K prefill", "16K prefill", "Chat decode (MTP)", "Agent decode (MTP)" ], "rows": [ [ "RTX 5090 D (24 GB, a China-market card)", "Q3_K", "3236.7", "4576.5", "130.2", "144.6" ], [ "RTX 5090 D (24 GB, a China-market card)", "NVFP4", "2668.5", "4357.1", "124.9", "135.5" ], [ "RTX 4090 24 GB", "Q3_K", "2637.5", "3551.1", "95.0", "104.9" ], [ "RTX Spark N1X 60 GiB", "Q3_K", "987.0", "1081.7", "47.2", "68.4" ], [ "RTX Spark N1X 60 GiB", "NVFP4", "996.4", "1066.9", "45.1", "68.9" ], [ "Apple M2 Ultra (Metal)", "Q3_K", "867.6", "908.7", "60.4", "63.0" ], [ "The same RTX 5090 D, capped to 16 GiB VRAM + 20 GiB RAM", "Q3_K", "430", "778", "62.4", "73.3" ], [ "The same RTX 5090 D, capped to 16 GiB VRAM + 20 GiB RAM", "NVFP4", "361.1", "619.7", "51.3", "59.9" ] ], "notes": [ "RTX 5090 D (24 GB) is the China-market GeForce RTX 5090 D v2, not the 32 GB RTX 5090 sold in the US. US readers can compare with the RTX 4090 row.", "\"RTX Spark N1X 60 GiB\" is Edge0's label. We have no hardware entry that matches it.", "Capped rows: Edge0 capped VRAM with --vram-reserve-mib 8192 and system RAM with a cgroup MemoryMax=20G, no swap. Edge0 says: \"These are test limits, not minimum system requirements.\"" ], "sources": [ { "label": "Model card", "url": "https://huggingface.co/Edge0/Edge0-120B-q3k" }, { "label": "Edge0 GitHub (docs/models/edge0-120b.md)", "url": "https://github.com/Edge0-AI/Edge0" } ], "checked": "2026-11-14" }, { "id": "vendor-benchmarks", "title": "Benchmark scores (Edge0's internal evaluations)", "text": "Edge0's own table, from its model card and GitHub docs. Edge0 says the Edge0, Qwen, DeepSeek and Claude results all come from its internal evaluations, and that the setup and reproduction steps will come with a technical report that is not out yet. We have not checked these scores. Edge0's blog lists different 120B figures (DeepSWE 53.6, Terminal-Bench 81.3); this table uses the card's.", "columns": [ "Model", "DeepSWE 1.1", "Terminal-Bench 2.1", "AA-LCR", "IFBench", "GPQA-Diamond", "HLE" ], "rows": [ [ "Edge0-145B", "64.6", "86.5", "79.3", "80.6", "90.4", "35.4" ], [ "Edge0-120B", "51.3", "75.3", "81.0", "79.3", "87.4", "31.0" ], [ "DeepSeek-V4-Flash-0731", "54.4", "82.7", "79.7", "79.2", "90.8", "33.8" ], [ "Qwen3.8-Flash-Next", "58.7", "83.3", "79.1", "81.3", "91.7", "35.9" ], [ "Qwen3.8-27B", "42.2", "73.0", "82.0", "79.5", "89.2", "30.8" ], [ "Claude-Opus-4.6 (Max)", "—", "78.2", "78.0", "62.5", "91.3", "40.0" ] ], "notes": [], "sources": [ { "label": "Model card", "url": "https://huggingface.co/Edge0/Edge0-120B-q3k" }, { "label": "Edge0 blog", "url": "https://edge0.ai/blog/introducing-edge0-145b-frontier-ai-capabilities-on-your-own-computer" } ], "checked": "2026-11-14" } ], "checks": [ { "kind": "official", "label": "Model card", "text": "Edge0's card: 120B total parameters and 8.64B active per token. The same 48 layers as Edge0-145B (36 linear-attention, an indexed full-attention layer every 4th), 512 routed experts per layer plus 1 shared, 10 active per token. The expert pool is 60.4B parameters at a latent width of 1280. It uses the same 51.2B fp8 n-gram embedding table (PLE), memory-mapped from the SSD, and the same 1.26B MTP draft head. The 51.2B table counts inside the 120B total. Context 256K. Edge0 gives the expert-weight builds as Q3_K 27.98 GiB and NVFP4 32.96 GiB; those are not the whole download. Edge0 describes the models as \"trained natively at 4 bits\" and derived from Qwen3.8-Flash-Next; its release-scope.json says \"No training-source commit or independent ownership audit is asserted\".", "source": "https://huggingface.co/Edge0/Edge0-120B-q3k", "source_label": "Model card", "checked": "2026-11-14" }, { "kind": "official", "label": "Weight files", "text": "Edge0/Edge0-120B-q3k totals 92.28 GB (ModelScope: 92,281,883,893 bytes) and Edge0/Edge0-120B-NVFP4 totals 97.63 GB (97,629,769,542 bytes on ModelScope; the Hugging Face sum is also 97.63 GB). experts.bin is 28,626,124,800 bytes in Q3_K and 33,974,255,616 in NVFP4; both include the same ple-fp8.gguf (51,200,246,112 bytes). NVFP4 latent_dense_q8.bin is 2,673,868,800 bytes and latent_dense_q4.bin 1,415,577,600. These are downloads, not the memory the engine needs.", "source": "https://huggingface.co/api/models/Edge0/Edge0-120B-q3k?blobs=true", "source_label": "Hugging Face API", "checked": "2026-11-14" }, { "kind": "official", "label": "Requirements (Edge0's docs)", "text": "Edge0's docs: \"Linux or Windows with an NVIDIA GPU and CUDA 13, or macOS on Apple Silicon with Metal. The macOS measurements were taken on an Apple M2 Ultra.\" and \"On Linux and Windows, at least 16 GiB of GPU memory and 20 GiB of system memory.\" The same docs also say the constrained-memory speed rows are measured test budgets, not general minimum requirements. Building edge0-strata needs the CUDA 13 toolkit (nvcc), and Visual Studio 2022 on Windows x64 for CUDA 13.0 to 13.2; Strata's install doc says CUDA 13 packages need NVIDIA driver 580 or newer. On a Mac only the Q3_K folder runs (NVFP4 is CUDA only), the Metal template sets a 32,768-token context, and the first start writes mtp-q2_0.gguf into the model folder, so the folder must be writable. The Edge0 Chat 0.2.0 app (Nov 11, 2026) has macos-arm64, windows-x64 and windows-arm64 builds.", "source": "https://github.com/Edge0-AI/Edge0", "source_label": "Edge0 GitHub (edge0-infer/docs/edge0-120b.md)", "checked": "2026-11-14" }, { "kind": "official", "label": "AMD", "text": "Edge0's engine has no AMD support. The edge0-strata fork removed upstream Strata's AMD/HIP, Intel SYCL and Docker setups (its EDGE0.md: \"Removed because neither the Linux CUDA build nor the server uses them\"), and the build instructions cover only NVIDIA CUDA 13 and Apple Metal. So a Radeon card such as the RX 7800 XT can't run it with Edge0's engine.", "source": "https://github.com/Edge0-AI/Edge0", "source_label": "Edge0 GitHub (edge0-strata/EDGE0.md)", "checked": "2026-11-14" }, { "kind": "official", "label": "Speculative decoding in Edge0's configs", "text": "Every shipped config uses MTP speculative decoding. The model.json in each repo sets --spec 4 --spec-min-p 0.5 --mtp ./mtp, the CUDA template (edge0-120b.json) sets --spec 4 with --mtp, the Metal template (edge0-120b-metal.json) also sets --spec 4, and edge0-infer's docs list mtp/ (\"Draft layer for speculative decoding\") as a required file. The bench config (model-bench.json in the 145B Q3_K repo) also sets --spec 4 with --mtp. Edge0's blog ties its \"80+ tokens/s\" figure to speculative decoding. Edge0 publishes no plain-decoding number, and this site does not estimate one.", "source": "https://huggingface.co/Edge0/Edge0-120B-q3k", "source_label": "Hugging Face repo (model.json)", "checked": "2026-11-14" }, { "kind": "checked", "label": "Which RTX 5090 D", "text": "Edge0's table names an \"RTX 5090 D 24 GB\". NVIDIA's China site lists the GeForce RTX 5090 D v2 with 24 GB of GDDR7 on a 384-bit bus. It is a China-market card, not sold in the US, and not the 32 GB GeForce RTX 5090 on this site, so Edge0's numbers are not attached to that page. US readers can compare with Edge0's RTX 4090 rows.", "source": "https://www.nvidia.cn/geforce/graphics-cards/50-series/rtx-5090-d-v2/", "source_label": "NVIDIA China, RTX 5090 D v2", "checked": "2026-11-14" }, { "kind": "vendor", "label": "Edge0's headline claims", "text": "Edge0's launch blog is about Edge0-145B: it says \"29GB peak memory\", that it \"runs on a compatible personal laptop, such as a MacBook Pro\", and \"80+ tokens/s decode\", without naming the hardware. For the 120B, Edge0's own residency figures on a 24 GB card are 22.93 GiB VRAM + 18.40 GiB RAM (Q3_K, about 41 GiB together) and 23.02 GiB + 23.40 GiB (NVFP4, about 46 GiB). Plan from those card figures, not a headline. These are Edge0's claims; we have not reproduced them.", "source": "https://edge0.ai/blog/introducing-edge0-145b-frontier-ai-capabilities-on-your-own-computer", "source_label": "Edge0 blog", "checked": "2026-11-14" }, { "kind": "official", "label": "Where Edge0's own figures disagree", "text": "Each repo's model.json says \"active_params\": \"6.1 B\", while the card and GitHub docs say 8.64B; this page uses the card's 8.64B. The card says 256K context, but the shipped CUDA template (edge0-120b.json) sets --max-context 131072 (128K) and the macOS Metal template (edge0-120b-metal.json) sets 32768; Edge0's blog says \"128K–256K context\". Edge0's blog table lists Edge0-120B at DeepSWE 53.6 and Terminal-Bench 81.3, while the card and GitHub docs list 51.3 and 75.3. This page shows the card figures.", "source": "https://github.com/Edge0-AI/Edge0", "source_label": "Edge0 GitHub and model.json", "checked": "2026-11-14" }, { "kind": "official", "label": "License", "text": "Two licenses apply to the weights: the Edge0 Community Model License 1.0 (dated 12 Nov 2026) and, on top, the Qwen Community License 1.0 of the base model (shipped as third-party/Qwen-LICENSE.txt). Free under Edge0's license: personal use and local deployment; research, education, benchmarking and publishing results; modifying, fine-tuning, quantizing, converting and non-commercial sharing; internal use inside a business; paid work by individuals that doesn't give customers access to the model; and commercial use of the outputs. You need a separate written license from Edge0 (info@edge0.ai) to sell or rent the model, weights, adapters or paid bundles built around it; to run a paid model service (an API or hosted endpoint) for third parties; or to sell an independent AI coding or office-assistant product built on it, even one that runs on the user's device. The Qwen license adds its own terms: a Model-as-a-Service or AI Work Assistant business needs a separate commercial license from Qwen, and a product with more than 100 million monthly active users or more than US$20 million in monthly revenue must show the model name. Edge0's permission does not replace Qwen's. Edge0's code (edge0-infer and the rest of the framework) is Apache 2.0, and edge0-strata keeps Strata's MIT license. This is a summary of the licenses, not legal advice.", "source": "https://huggingface.co/Edge0/Edge0-120B-q3k", "source_label": "Model card", "checked": "2026-11-14" }, { "kind": "community", "label": "Stock llama.cpp", "text": "The Henry Labs' RTX 3090 report says: \"Stock llama.cpp cannot load this model: the pack keeps its experts outside the GGUF in a latent layout that upstream's qwen4exp loader does not support.\" That report is about the 145B Q3_K pack; the 120B uses the same pack layout.", "source": "https://github.com/thehenrylabsdev/edge0-145b-rtx3090-benchmark/blob/main/REPORT.md", "source_label": "The Henry Labs REPORT.md", "checked": "2026-11-14" }, { "kind": "checked", "label": "Ollama and MLX", "text": "Ollama's library pages /library/edge0, /library/edge0-145b and /library/edge0-120b return 404. We found no MLX conversion.", "source": "https://ollama.com/library/edge0", "source_label": "Ollama library", "checked": "2026-11-14" }, { "kind": "unknown", "label": "Not documented", "text": "Unknown, and not estimated here: a plain-decoding (no speculation) speed, a minimum SSD speed, the hardware behind Edge0's 29 GB headline, Edge0's evaluation setup (promised with a technical report that is not out yet), speeds for long outputs, and any fit or speed on a PC or Mac in our catalog.", "source": "https://edge0.ai/blog/introducing-edge0-145b-frontier-ai-capabilities-on-your-own-computer", "source_label": "Not in Edge0's blog or card", "checked": "2026-11-14" } ], "related": [ { "href": "/models/edge0-145b/", "k": "Model", "t": "Edge0-145B, the larger Edge0 model on the same engine" }, { "href": "/models/qwen3.8-flash-next/", "k": "Model", "t": "Qwen3.8-Flash-Next, the base model Edge0 derived it from" }, { "href": "/guides/dgx-spark-cluster-two-sparks/#ssd-streamed-setups-on-one-spark", "k": "Guide", "t": "How Strata-style SSD streaming works and what we've seen on a Spark (Edge0's engine is a Strata fork)" }, { "href": "/guides/hugging-face-alternatives-download-models/", "k": "Guide", "t": "Other places to download models besides Hugging Face" } ], "alt_checked": "2026-11-14", "alt_downloads": [ { "source": "modelscope", "url": "https://modelscope.cn/models/Edge0/Edge0-120B-q3k", "uploader": "official", "note": "Q3_K build" }, { "source": "modelscope", "url": "https://modelscope.cn/models/Edge0/Edge0-120B-NVFP4", "uploader": "official", "note": "NVFP4 build" } ] } ]