Recursive launches Flint, a laptop coding model it says was largely built by AI | TechCrunch

Recursive, a startup founded last year with $650 million in funding at a $4.65 billion valuation, released Flint on a Wednesday in October. It's a free coding model that runs on a laptop and was built largely by the company's own AI research system, the Eureka Machine. According to Recursive, the Eureka Machine helped with nearly every step, from pretraining to post-training and inference, letting a small team go from nothing to a finished model in under four months. The model is on Hugging Face under an Apache 2.0 license, so anyone can download, use, and modify it. CEO Richard Socher, formerly Salesforce's chief scientist, is leading the effort. Flint is a mixture-of-experts model with 40 billion parameters, but it activates only 2.5 billion for any given task, which keeps it light enough for an ordinary machine. That matters for developers who want offline use, private code, and no per-token cloud fees. In Recursive's tests on a MacBook Pro with an M4 Pro chip, it produced about 131 tokens per second, roughly double what comparable models like Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning managed on the same hardware. On SWE-Bench Verified, which tests fixing real software bugs, it beat other models under 50 billion parameters. Wired's Will Knight found it handled his small coding requests reliably where other small models stumbled, though the company itself admits it needs more hand-holding than large paid assistants and works best when big tasks are broken into smaller steps. The bigger story is the method. Recursive is betting on recursive self-improvement, where AI helps build better AI, and Flint is its first public test. The Eureka Machine generates ideas, runs experiments, and uses the results to pick the next step. Its concrete contributions, per Socher, included a sparser model architecture, low-bit training that shrank Flint by 71% with little quality loss, and custom kernels for faster computation. Recursive hasn't said how much researcher time this saved, and the current setup is still an AI assistant under human supervision. Socher's stated goal is full self-improvement with humans able to halt it, a prospect some researchers find worrying. The company tied its $410 million Amazon Web Services compute deal to the idea of "agent count" over headcount, and OpenAI has reported similar gains, with its agents logging 3.1 workdays for each human researcher day by mid-August.