| Official | Model card | 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). | Model card, checked Nov 11, 2026 |
| Official | Weight files | 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. | Hugging Face API, checked Nov 11, 2026 |
| Official | Architecture config | 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. | config.json, checked Nov 11, 2026 |
| Our check | Same config as Qwen3.5-122B-A10B | 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. | Qwen config.json, checked Nov 11, 2026 |
| Official | MLX for Macs | 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. | MLX 4-bit repo, checked Nov 11, 2026 |
| Community claim | ezoushen MLX mixed 4.46 bpw | 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. | ezoushen card, checked Nov 11, 2026 |
| Official | Single-GPU serving claim | 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. | Model card, checked Nov 11, 2026 |
| Official | FP8 checkpoint | ornith-ai/Ornith-1.5-122B-FP8 is 132,949,613,832 bytes (132.95 GB). A download size, not memory. | FP8 repo, checked Nov 11, 2026 |
| Official | NVFP4 checkpoint | ornith-ai/Ornith-1.5-122B-NVFP4 is 77,627,904,940 bytes (77.63 GB). A download size, not memory. | NVFP4 repo, checked Nov 11, 2026 |
| Community claim | llama.cpp | 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. | bartowski GGUF (commit a8bc08fa), checked Nov 11, 2026 |