Recursive releases Flint, a 39.8B coding-agent model under Apache 2.0
Original title: RecursiveSI/flint — new model trending #29 on Hugging Face
About 2.5 billion parameters are active per token; Recursive reports 54.9% on Terminal-Bench 2.1 and 89.7% on SWE-bench Pro v2.
Recursive published Flint, a mixture-of-experts language model aimed at coding agents, on Hugging Face with a BF16 checkpoint and an SGLang serving kit. It has 39.8 billion total parameters, roughly 2.5 billion active per token, text-only input and output, tool use, and a stated 262,144-token context. Recursive’s own evaluations use temperature 1.0 and mark some comparison scores as developer-reported. The weights are Apache 2.0; hosted API terms are separate.
Unconfirmed: Benchmark figures are Recursive’s release evaluations; some comparison scores are developer-reported and setups differ.
- Architecture is 49 layers, with 256 routed experts and 6 selected per token.
- Stated knowledge cutoff is June 2026.
- One B300 or H200 GPU per replica is the documented serving setup; the download is about 80 GB.
- Quantized NVIDIA and Apple Silicon variants are prepared separately.
- Safety fine-tuning is included, but the card says harmful output and unsafe tool calls remain possible.
From the original article
4 min read at huggingface.cotext-generation · 0 downloads · 159 likes · license apache-2. 0
By RecursiveSI · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.