Recursive Ships Flint, a Free 40B Coding Model Its Own AI Helped Build in Under Four Months

Recursive, the year-old startup trying to build AI that improves itself, has released its first model: Flint, a free coding model small enough to run on an ordinary laptop and, by the company's own account, built mostly by AI. Flint went up on Hugging Face on Wednesday under the permissive Apache 2.0 license, and Recursive says its automated research system, the Eureka Machine, contributed to nearly every stage of development — from the model's shape to its training recipe to the custom kernels that make it fast. A small team took the model from scratch to release in under four months.

Flint is a mixture-of-experts model with 40 billion parameters, of which only about 2.5 billion activate for any given task. That frugality is what lets it live on a laptop: it runs locally, works offline, keeps code on the user's own machine, and carries no per-token cloud fees. The contrast with the current open-weight trend is stark. Meta's Muse Spark Open, released last week, uses 816 billion parameters and needs at least a terabyte of memory; Flint targets the opposite end of the market.

The speed numbers are the headline claim. Recursive reports roughly 131 tokens per second on a MacBook Pro with Apple's M4 Pro chip — about double the 58 to 69 tokens per second it measured for comparable models on the same hardware, including Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning. On Nvidia's RTX Spark the company says Flint tops 150 tokens per second, and Nvidia's own conference materials put it above 220 tokens per second on a 64GB RTX Spark configuration. All of these figures are company-reported and have not been independently verified.

On benchmarks, Recursive says Flint outperformed every model under 50 billion parameters on SWE-Bench Verified, a test of real-world bug fixing, and that on its own six-test average only Alibaba's Qwen3.8 27B scored higher. Early hands-on impressions were favorable: Wired's Will Knight used the model on a 48GB Mac to build a to-do app and a web research tool and found it handled his coding requests without the stumbles he usually sees from small local models. The company concedes the caveat itself: Flint needs more guidance than paid cloud assistants, and large tasks should be broken into smaller steps.

The more interesting story is who built it. The Eureka Machine comes up with research ideas, runs experiments, and uses the results to decide what to try next. Recursive credits it with a sparser model architecture, a low-bit training scheme that shrank Flint by 71 percent with almost no quality loss, and custom kernels that added 55 percent throughput on RTX Spark without touching the weights. The company did not say how much researcher time the system saved. The long-term pitch is recursive self-improvement under human oversight — each model makes the Eureka Machine a better researcher, which builds the next model faster. OpenAI offered a comparable data point last month: by mid-August its agents were logging 3.1 workdays of work for every one worked by a human researcher.

Recursive emerged from stealth in May with 650 Million Dollars in funding at a 4.65 Billion Dollar valuation, backed by GV, Nvidia and AMD Ventures, and is led by Richard Socher, Salesforce's former chief scientist. In June it signed a 410 Million Dollar compute deal with Amazon Web Services that the company described as being less about headcount and more about agent count.

Flint is best read as an early, controlled experiment in a debate that usually stays theoretical. If the Eureka Machine genuinely compressed a model development cycle into four months, the measurable test is the next release: whether the gap between models keeps shrinking on an accelerating curve, or whether human researchers remain the bottleneck that AI assistance only partly relieves. Either result would be informative — and Recursive has just made both of them public.