| Official | Model card | 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). | Model card, checked Nov 13, 2026 |
| Official | License | Apache 2.0, in the repo's LICENSE file. This is a summary of the license, not legal advice. | LICENSE, checked Nov 13, 2026 |
| Official | Weight files | 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. | Hugging Face API, checked Nov 13, 2026 |
| Official | NVFP4 checkpoint (experimental) | 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. | NVFP4 repo, checked Nov 13, 2026 |
| Vendor claim | DGX Spark (Ricursive's claim) | 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. | Model card, checked Nov 13, 2026 |
| Vendor claim | Benchmark scores (Ricursive's) | 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. | Model card, checked Nov 13, 2026 |
| Official | DFlash2 draft model | 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. | DFlash2 repo, checked Nov 13, 2026 |
| Our check | DFlash2 in llama.cpp (not merged) | 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. | llama.cpp PR #31904, checked Nov 13, 2026 |
| Community claim | GGUF files (AtomicChat) | 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. | AtomicChat GGUF card, checked Nov 13, 2026 |
| Community claim | GH200 run (AtomicChat's figure) | 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. | AtomicChat GGUF card, checked Nov 13, 2026 |
| Our check | llama.cpp | 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. | llama.cpp PR #31692, checked Nov 13, 2026 |
| Community claim | Ollama and LM Studio | 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. | AtomicChat GGUF card, checked Nov 13, 2026 |
| Community claim | MLX (unofficial) | 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. | frogggias MLX card, checked Nov 13, 2026 |