Recursive Flint

Recursive Flint is an open-weight 40B mixture-of-experts model built for fast, accurate, local AI coding.

Highlights

• Agentic coding performance: delivers leading performance among MoE models under 50B parameters on benchmarks including SWE-bench Pro v2, Terminal-Bench 2.1, and SWE-bench Verified.
• Efficient sparse architecture: combines 40B total parameters with approximately 2.5B active parameters per token, significantly reducing the computation required for generation compared with similarly sized dense models.
• Long-context workflows: supports up to 262,144 tokens of context, including the prompt and generated output, for reasoning and multi-turn tool use.
• Open and customizable: available under the commercially permissive Apache 2.0 license, giving developers the freedom to run, customize, and deploy Flint locally.