Flint: Recursive's open-weight 40B MoE coding model, built largely by its Eureka Machine

Recursive Superintelligence released Flint, a free 40B-parameter mixture-of-experts coding model (2.5B active per token) that runs offline on a laptop, built in under four months with its automated Eureka Machine research system.

Why it matters

What you could build with it

Does it hold up?

Early evidence only: Recursive's own benchmarks put it at 74.2% on SWE-bench Verified, with the caveat that it needs more guidance than closed models; a Wired hands-on built small programs successfully on a 48GB Mac, but no independent benchmark verification exists yet.

Built with Flint

• Recursive Releases Flint, an Open-Weight 40B MoE Local Coding Model for Agents with 2.5B Active Paramsmarktechpost · Technical breakdown with benchmark tables: SWE-bench Verified 74.2%, Terminal-Bench 2.1 54.9%, 262K documented context vs 512K in the blog, runs on 48GB Apple Silicon or NVIDIA GPUs.
• Recursive Flint: Richard Socher's AI-built open coding modelmadrobot.blog · Independent writeup noting Flint's benchmarks are company-run with no independent verification, and rival figures in its charts were run under different setups.
• $650M-Backed Recursive Says AI Helped Build Its First Modelfourweekmba · Reports Socher's X launch post (Oct 21): 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark, free to try on opencode and pidotdev.

Learn more

technical deep dive →

First spotted on article: source.

More AI models