A coding AI built largely by another AI now fits on your laptop
Startup Recursive says its automated research system helped build Flint, a free, fast coding model that runs offline on an ordinary laptop.
What happened: Recursive, a one-year-old startup chasing self-improving AI, released Flint on Wednesday: a free coding model light enough to run on a laptop instead of in the cloud. The company says its automated research system, called the Eureka Machine, helped design nearly every part of Flint — from training to final tuning — letting a small team finish the model in under four months. Flint is a mixture-of-experts model with 40 billion total settings but only taps 2.5 billion at a time, keeping it light enough for ordinary hardware. It's free to download on Hugging Face under the open Apache 2.0 license.
Why it matters: Flint is a real-world test of "recursive self-improvement" — AI systems building better AI, which build even better AI after that — an idea that excites researchers chasing faster progress and worries others who fear it could eventually slip past human control. For now, Recursive keeps people in charge: the Eureka Machine proposes and tests ideas, but humans watch and can halt it. The stakes go beyond one lab — OpenAI recently said its research agents were logging more than three workdays of output for every day a human researcher puts in. Meanwhile, as more code gets written by AI, companies like Blacksmith, which just hit a $1 billion valuation testing that code, are racing to build the plumbing needed to check whether any of it can be trusted.
How it works, plainly: Because Flint runs on a user's own machine rather than a company's servers, developers can code offline, keep their work private, and skip the per-use fees cloud AI tools charge. The Eureka Machine's contributions reportedly included a leaner model design, a compression technique that shrank Flint by 71% with little quality loss, and custom software that speeds up its computations. In Recursive's own tests, Flint generated about 131 tokens per second on a MacBook Pro with an M4 Pro chip — roughly double similarly sized models from Google and Nvidia — and topped 150 tokens per second on Nvidia's RTX Spark hardware. On a bug-fixing benchmark called SWE-Bench Verified, it beat every other model under 50 billion parameters that Recursive tested.
The rollout: Reviews have been cautiously positive: Wired's Will Knight said Flint handled every small coding task he gave it, though bigger paid rivals like Claude or Codex still win on harder jobs, and Recursive itself says Flint needs tasks broken into smaller steps to perform well. The company, led by former Salesforce chief scientist Richard Socher, emerged from stealth in May with $650 million in funding at a $4.65 billion valuation from backers including GV, Nvidia, and AMD Ventures, and later signed a $410 million compute deal with Amazon described as being "more about agent count" than headcount. Recursive says each new model should sharpen its research system, speeding up the next release.