- Startup Recursive released Flint, a free coding model small enough to run on a laptop, built largely by the company’s own “Eureka Machine” AI research system.
- Flint is a mixture-of-experts model using 2.5 billion of its 40 billion parameters per task, making it far smaller than Meta’s 816-billion-parameter Muse Spark Open.
- On a MacBook Pro with an M4 Pro chip, Flint generated about 131 tokens per second, roughly twice the 58-69 tokens per second of comparable models, including Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning.
- Flint outperformed other sub-50-billion-parameter models on SWE-Bench Verified, a benchmark that measures real-world software bug fixing, and is free to download on Hugging Face under the Apache 2.0 license.
What Happened
Recursive released its first model, Flint, on October 21, 2026, saying its automated Eureka Machine research system helped with nearly every step of development, letting a small team build it from scratch in under four months, TechCrunch reported.
Why It Matters
Flint is an early public test of recursive self-improvement — AI building better AI — one of the industry’s most debated ideas. Recursive, led by CEO and Salesforce’s former chief scientist Richard Socher, emerged from stealth in May with $650 million in funding at a $4.65 billion valuation from GV, Nvidia, and AMD Ventures.
Technical Details
Because Flint runs locally rather than on cloud servers, developers can use it offline, keep code private, and avoid per-token cloud fees. On Nvidia’s RTX Spark hardware, Flint topped 150 tokens per second, and Recursive says the model beat other sub-50-billion-parameter systems on SWE-Bench Verified.
Who’s Affected
Developers wanting local, private coding assistance get a new lightweight option competing with Meta’s much larger Muse Spark Open, Google’s Gemma 4, and Nvidia’s Nemotron 3.5 Lightning. Recursive’s approach also puts its self-improving research methodology under public scrutiny for the first time.
What’s Next
Recursive describes Flint as a limited, human-overseen version of its self-improvement ambitions rather than a full release of autonomous AI research capability, and has not detailed a timeline for larger models built the same way.
