Oriel — AI & technology worth noticing, every day.

RecursiveSI/flint

AI model · https://huggingface.co/RecursiveSI/flint

First seen by Oriel on .

AI summary

Recursive Flint is a 39.8B-parameter mixture-of-experts coding model, with about 2.5B active parameters per token, built for repository investigation and software engineering agents.

  • The model supports text, reasoning, and multi-turn tool use, and reports 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2, and 74.2% on SWE-bench Verified.
  • Its 49 layers include 47 MoE layers, each selecting 6 of 256 routed experts alongside an always-active shared expert; the documented context configuration is 262,144 tokens.
  • The repository includes a BF16 checkpoint, tokenizer, chat template, and SGLang serving kit; deployment requires one B300 or H200 GPU per replica, and Transformers modeling code is not included yet.
  • Flint was trained from scratch using pretraining, mid-training, supervised fine-tuning, and reinforcement learning, including text, code, instruction-following, and agentic tasks.
  • The Apache 2.0-licensed model can produce faulty code or unsafe tool calls; users should test code, enforce tool permissions, and evaluate the specific checkpoint and runtime.

#mixture-of-experts #coding agents #sglang #language models

Generated by gpt-5.6-sol. Check the original for details.

Metric trends

Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained

Likes

Total likes reported by the source.

  • Hugging Face Models: 1,350, latest observation ; +1,237 over observed period since

Downloads (30d)

Rolling 30-day downloads. A change in this value is not new downloads during the observed period.

  • Hugging Face Models: 38,357, latest observation ; +38,357 over observed period since

Aliases

SourceExternal IDURL
Hugging Face ModelsRecursiveSI/flinthttps://huggingface.co/RecursiveSI/flint Read discussion