Recursive releases Flint 40B coding agent model on Hugging Face
Recursive announced the open‑source release of Flint, a mixture‑of‑experts language model built for coding agents. The model has 39.8 billion total parameters but only about 2.5 billion active parameters per token, following the “40B‑A2.5B” size label. Flint is available as a BF16 checkpoint on Hugging Face together with a tokenizer, chat template and an SGLang serving kit, and separate…
Key points
- Flint totals 39.8 B parameters with roughly 2.5 B active per token (40B‑A2.5B designation).
- It scores 89.7 % on SWE‑bench Pro v2 and 74.2 % on SWE‑bench Verified in Recursive’s tests.
- Flint runs on a single B300 or H200 GPU using the SGLang serving kit, with NVIDIA and Apple Silicon quantized variants.
In Recursive’s own evaluations Flint achieved 54.9 % on Terminal‑Bench 2.1, 89.7 % on SWE‑bench Pro v2 and 74.2 % on SWE‑bench Verified, using temperature 1.0, top‑p 0.95 and top‑k 20. The model runs on a single B300 or H200 GPU per replica via the SGLang kit, and supports OpenAI‑compatible tool calls. Flint is released under the Apache 2.0 license, while hosted API terms and data practices are separate. The release notes warn that the model can still generate incorrect or harmful code and that users should enforce safety checks and execution boundaries.
RecursiveSI/flint (40B A2.5B) · Hugging Face
huggingface.co · 21 October 2026
Loading the full article…
This text was published by huggingface.co. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
3sources- Reddit discussionreddit.com
- Reddit discussionreddit.com
- Recursive releases FlintPress · recursive.com ·
- This Startup Is All In on Recursive Self-Improvement. Its Latest Model Is Scary GoodPress · Wired AI ·
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- Dharma AI says VocabX reduces inference cost by 77% for Qwen3-4B grader · 1 src
- Anthropic cuts Haiku 5.5 API costs by 90% for short inputs · 4 src
- OpenAI says its AI agents weakened internal misalignment monitoring system · 24 src
- IBM releases Granite Speech 5.0 2B for multilingual speech recognition · 1 src
- InclusionAI releases Ming-3.0-audio-flash, a multilingual audio-language model · 3 src
Comments
via GitHub Discussions