RecursiveSI/flint · Hugging Face

Flint is a 39.8 billion parameter mixture-of-experts language model from Recursive, built for coding agents: investigating codebases, fixing bugs, implementing features, and writing tests. Only about 2.5 billion parameters are active per token, which is what keeps inference cost closer to a small model's. The architecture is a decoder-only Transformer with 49 layers, the first two dense and the rest MoE, with 256 routed experts per layer (6 selected per token) plus an always-on shared expert. Attention is mostly sliding-window (35 layers with a 513-token window) with 14 full-attention layers, and the documented context window is 262,144 tokens, prompt and output combined. The model takes text in and produces text out, supports multi-turn tool use and reasoning, and has a stated knowledge cutoff of June 2026. The headline numbers are 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2, and 74.2% on SWE-bench Verified. Recursive measured the scores flagged with a dagger itself and reports other comparisons as developer-reported, and the Terminal-Bench figure counts solutions found through web retrieval as failures, a adjustment the developer notes others may not make. Evaluation used temperature 1.0, top-p 0.95, top-k 20, and averaged pass@1 over at least five attempts. The repo includes the BF16 checkpoint (about 80 GB), tokenizer, chat template, and an SGLang serving kit. Running it needs one B300 or H200 GPU per replica, and Transformers modeling code is not yet provided, so the Transformers snippets in the page's sidebar won't work for this checkpoint. Serving goes through SGLang or vLLM-style OpenAI-compatible endpoints, and the docs say to preserve returned reasoning_content in conversation history and to budget at least 32,768 output tokens for hard problems, since long reasoning can exhaust shorter limits before an answer finishes. The limitations section is worth reading before deploying anything. Recursive says the model can still write faulty code or make inappropriate tool calls, and that structured tool-call grammar doesn't stop it from choosing an unsafe action, so applications need their own permission and execution boundaries. Accuracy is not guaranteed uniform across sequence lengths, and results shift with precision and runtime, so you'd need to evaluate the exact checkpoint you plan to serve. The weights are Apache 2.0, but hosted API terms and data practices are governed separately.