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

Recursive, the startup led by former Salesforce chief scientist Richard Socher, released Flint, a free coding model under Apache 2.0 that runs locally. The company says its automated research system, the Eureka Machine, contributed to nearly every stage of building it, letting a small team go from scratch to release in under four months.

Two stories ride together here. One is practical: a fast, private, offline coding model with no usage fees. The other is strategic: an early public test of recursive self-improvement, one of the most debated ideas in AI.

Let's take the useful part first, because it's real. Flint is a mixture-of-experts model that activates about 2.5 billion of its 40 billion parameters at a time. In Recursive's own tests it ran around 131 tokens per second on an M4 Pro MacBook Pro, roughly double similar models on the same machine. It beat other sub-50-billion-parameter models on SWE-Bench Verified, and an early hands-on review at Wired came back positive. A capable coding assistant that keeps code on your own machine and costs nothing per token is a genuine win for developers, students and anyone handling sensitive code.

The caveats: most of the performance numbers are Recursive's. The company itself says Flint needs more guidance than the big paid assistants and suggests breaking tasks into smaller steps. That's honest, and it frames Flint as a strong local tool, not a replacement for frontier cloud models.

The bigger question is the Eureka Machine. Recursive credits it with a sparser model shape, low-bit training that cut Flint's size by 71 percent with little quality loss, and custom kernels. It did not say how much researcher time that saved. So we know AI contributed, but not how much of the judgment was human. That distinction is the whole debate.

I'm watching this closely because the trajectory is explicit. The stated long-term goal is full recursive self-improvement, with humans overseeing and able to stop it. Today it's a supervised research assistant producing a laptop model, which is about as low-risk as this gets. But the company's own framing, that each model should make the system a better researcher and build the next one faster, describes a loop where oversight has to scale as fast as capability. Nobody has shown how to guarantee that yet. We'd like Recursive to publish what its stop mechanisms actually look like before the models get bigger.

Recursive emerged from stealth in May with $650 million at a $4.65 billion valuation, from backers including GV, Nvidia and AMD Ventures, and signed a $410 million compute deal with AWS in June. The trend of AI doing AI research is broader: OpenAI said last month that by mid-August its agents were logging 3.1 workdays for every one worked by a human researcher.