Flint is a mixture-of-experts model, meaning it activates only 2.5 billion of its 40 billion parameters for any given task. That keeps it light enough for an ordinary laptop and far smaller than Meta's Muse Spark Open, an open model released last week that uses 816 billion parameters and needs at least a terabyte of memory. In Recursive's tests, Flint generated about 131 tokens per second on a MacBook Pro with an M4 Pro chip, roughly twice the 58 to 69 tokens per second managed there by Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning, and topped 150 tokens per second on Nvidia's RTX Spark.

Because it runs locally rather than in the cloud, developers can use it offline, keep code private and avoid usage fees. On SWE-Bench Verified, a benchmark measuring how well AI fixes real-world software bugs, Recursive says Flint outperformed other models under 50 billion parameters, and only Alibaba's Qwen3.8 27B scored higher on the company's own six-test average. The company concedes Flint needs more guidance than big paid coding assistants, though Wired's Will Knight reported that it handled his small-program builds, including a to-do list app, nailing his coding requests where smaller models often stumble.

Flint is an early public test of recursive self-improvement, the idea that AI can build better AI, and one of the industry's most debated concepts. Recursive frames the Eureka Machine as a human-supervised research assistant rather than an autonomous system, and Socher told Wired the long-term goal is full recursive self-improvement with humans able to stop the process. The company emerged from stealth in May with $650 million in funding at a $4.65 billion valuation, with backers including GV, Nvidia and AMD Ventures.