Recursive Launches Flint, Laptop Coding Model Built Largely by AI

Recursive, an AI startup founded last year, released Flint on Wednesday, a free coding model designed to run on laptops and developed largely by the company’s automated research system, the Eureka Machine. Recursive says Eureka contributed to nearly every stage of Flint’s creation, from pre-training and post-training to inference, enabling a small team to build the model from scratch in under four months. The development offers an early public test of recursive self-improvement, the contested idea that AI systems can help build better AI systems under human oversight. Flint is designed for local use, allowing developers to work offline, keep code private, and avoid the usage fees charged by cloud-based AI services. The model is available for free on Hugging Face under the permissive Apache 2.0 license. It is a mixture-of-experts model with 40 billion parameters, but only activates about 2.5 billion parameters at a time, making it compact enough for ordinary laptops. Recursive said this architecture, along with low-bit training that shrank the model by 71% with minimal quality loss and custom kernels for faster computation, was among the improvements identified by Eureka. In Recursive’s tests, Flint generated about 131 tokens per second on a MacBook Pro with Apple’s M4 Pro chip, roughly twice the speed of comparable models including Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning. On Nvidia’s RTX Spark, it exceeded 150 tokens per second. Flint also performed strongly on SWE-Bench Verified, a benchmark measuring real-world software bug fixing, outperforming other models under 50 billion parameters. Across six additional tests, only Alibaba’s Qwen3.8 27B scored higher than Flint. Early external testing was favorable, with Wired’s Will Knight reporting that the model reliably handled small coding tasks, though Recursive acknowledged that larger jobs may require users to break work into smaller steps. The launch advances Recursive’s broader strategy of using AI to accelerate AI research. The company emerged from stealth in May with $650 million in funding at a $4.65 billion valuation, backed by investors including GV, Nvidia, and AMD Ventures. It is led by CEO Richard Socher, formerly chief scientist at Salesforce. Recursive has also secured a $410 million compute deal with Amazon Web Services, which it described as focused less on hiring more researchers and more on deploying more AI agents. Recursive’s long-term ambition is for each new model to make the Eureka Machine a more effective researcher, helping it build the next model more quickly, with humans supervising and retaining the ability to stop the process. The idea remains controversial in AI circles because self-improving systems could eventually operate beyond human control. For now, Recursive’s public example remains narrowly scoped: a laptop-sized model produced with substantial machine assistance, but still under human direction.