The first model from the new company Ricursive Intelligence: a 65B parameter dense Transformer (no MoE), pre-trained from scratch on a 12T token corpus primarily consisting of English and code, supporting a maximum context length of 1M. This repository is the BF16 instruction-tuned version, which also releases the NVFP4 instruction-tuned version, the BF16 base version, and the DFlash2 draft model for speculative decoding. The authors state that they started from the Olmo 3 baseline and iteratively modified the training recipe; during development, the team grew from 1 to 5 people, and the architecture was changed to a "narrow and deep" structure similar to Gemma 3N (d_model 5120, 80 layers). The repository was created on 11 4, and it has not been included in the last 7 periods.
Upvotes 112 · Likes 114 · Downloads 0
Reasons
- The authors claim that agentic coding performance exceeds its size: Terminal-Bench 2.1 80.5, higher than Reflection Beam (80.1) at 501B, Muse Spark Open (76.4) at 816B, Inkling (63.8) at 975B, and Nemotron 3 Ultra (56.4) at 550B in the table; SWE-bench Pro v2.0 Hard 79.1 is also the highest in the table, with Math AIME 2025 98.3 and HMMT Feb 2026 90.5.
- Dense 65B deployment is simple: the authors claim it fits on local devices like the DGX Spark, for which they released the NVFP4 version (MLP uses NVFP4, attention retains BF16); on 2 B200 GPUs with batch 8, without quantization or custom kernels, it can reach 250 tokens per second.
- The training process is well-documented: tables list the number of tokens, steps, sequence lengths, and batch sizes for each stage, including pre-training, mid-training, long-context, SFT, long-context SFT, and RL; the base version is a checkpoint with only pre-training, without mid-training or synthetic QA data, making it suitable for post-training research.
Usage
- Deployment method provided by the authors (inference/server directory in the repository): Linux + CUDA Toolkit 13 + uv, SGLang 0.5.19, two B200 GPUs, run in that directory: CUDA_VISIBLE_DEVICES=0,1 uv run --script serve.py --model-path <this repository> --speculative-draft-model-path <DFlash2 draft model>; default TP=2, 1Mi context, service at 127.0.0.1:30000, OpenAI-compatible interface.
- Tool calling: pass tools and tool_choice: auto in OpenAI format; the response separates reasoning_content, content, and tool_calls; for subsequent turns, put the assistant message back as-is, then add a tool message with tool_call_id. To use tool_choice: required, set the function to strict=true and additionalProperties=false.
- The model includes custom code (RicursiveForCausalLM, requires trust_remote_code); the chat template has no thinking toggle or reasoning intensity levels, so max_tokens must be sufficient for both thinking and the answer.
Risks
- Preview version: the authors state that post-training is lighter and focused on agentic coding, with SciCode 39.4, DeepSWE 1.1 44.3, and MCP Atlas 71.67 scoring lower in the table. Scores are self-reported by the authors, italicized scores in the table are run by the authors for competitors, and the authors acknowledge that agentic evaluations are difficult to reproduce.
- Comparison is only with US open-source models, without comparisons to Qwen, DeepSeek, Kimi, GLM, etc.; the pre-training corpus is primarily English and code, text-only, so Chinese capability is unknown (base version C-Eval 69.6, lower than the two DeepSeek models in the table).
- The Colab mentioned in the card has no link provided; the card for the NVFP4 version only has a brief description, with no running steps on devices like DGX Spark; existing documentation only covers B200 + SGLang; this is the first model from a new company, and future maintenance status is unknown.