Flint 40B

Release Summary

Recursive's open-weight coding model for local agents. 39.8B total parameters, about 2.5B active per token, Apache 2.0 weights at RecursiveSI/flint, and no published API price. The model card documents a 262,144-token context shared by the prompt and the output. The announcement also describes windows up to 512k and recommends staying within 256k. On the card's BF16 chart, Terminal-Bench 2.1 is 54.9% with Terminus 2, SWE-bench Pro v2 is 89.7% with mini-swe-agent, and SWE-bench Verified is 74.2%. The NVFP4 checkpoint on that chart scores 53.7, 88.5, and 74.1. Qwen3.8 27B is higher on the same three panels, at 73.0, 94.9, and 75.6. Muse Glimmer 30B is 51.7, 86.0, and 76.0. The optimized engine reports more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 on a 64GB RTX Spark at short contexts. Knowledge cutoff is June 2026.

Timeline