🤖Recursive's Flint: 40B Coding Model Hits 131 tok/s on a Mac
TL;DR
Recursive released Flint, a free Apache 2.0 coding model with 40B total and 2.5B active parameters. It hits about 131 tokens per second on an M4 Pro MacBook, roughly twice similar models.
Recursive released Flint, a free Apache 2.0 coding model with 40B total and 2.5B active parameters. It hits about 131 tokens per second on an M4 Pro MacBook, roughly twice similar models. The startup says its own AI research system built most of it.
Key Points
Mixture-of-experts design: 40B parameters total, 2.5B active per task
About 131 tok/s on an M4 Pro MacBook Pro versus 58 to 69 for comparable models; over 150 on Nvidia RTX Spark
Beat other sub-50B models on SWE-Bench Verified; only Alibaba's Qwen3.8 27B scored higher on Recursive's six-test average
Low-bit training shrank the model 71% with almost no quality loss
Needs more hand-holding than large paid assistants, so break big tasks into steps
Why It Matters
A capable local coding model means private, offline code assistance on a laptop. The catch is that the benchmark averages come from Recursive itself, so test it on your repo.
Quick Facts
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,569 builders reading daily.