Recursive has released Flint, its first AI model. Flint is a free coding model that runs on a laptop, and the startup says its own automated research system helped build nearly all of it.
The release is an early public test of recursive self-improvement, the idea that AI systems can build better versions of themselves. It is one of the most debated ideas in the industry, and some researchers fear self-improving AI could slip out of human control.
What Flint is
Flint is a mixture-of-experts model with 40 billion total parameters. Only about 2.5 billion parameters activate for each token, which is what keeps it light enough for an ordinary laptop.
The weights are free to download on Hugging Face under the Apache 2.0 license, which allows free use, modification and commercial deployment. Developers can run Flint offline, keep their code private and avoid the usage fees that cloud AI services charge.
Recursive says Flint generates more than 125 tokens per second on a MacBook Pro with an M4 Pro chip, and more than 150 tokens per second on a 64GB NVIDIA RTX Spark. That is up to twice as fast as comparable models in the company’s tests.
How the Eureka Machine built it
Recursive’s automated research system, called the Eureka Machine, contributed to nearly every part of Flint’s development. It proposes ideas, implements them, tests the results and uses them to decide what to try next.
A small team of people and agents built the model from scratch in under four months. The Eureka Machine explored hardware-efficient architectures, searched for quantization recipes that shrink memory use and wrote custom kernels that speed up inference.
Recursive calls Flint a practical test of the system. Its long-term goal is recursive self-improvement, where each new model makes the Eureka Machine a better researcher and helps build the next generation faster.
Performance and early tests
In Recursive’s own tests, Flint leads mixture-of-experts models with fewer than 50 billion parameters on agentic coding benchmarks, including SWE-Bench Pro v2 and SWE-Bench Verified. These company-run benchmarks have not been checked independently.
The company acknowledged that Flint may need more guidance than larger closed coding models and suggested breaking complex tasks into smaller steps. An early independent test found it handled a series of coding requests on a Mac without errors.
The company behind Flint
Recursive was founded last year and is led by CEO Richard Socher, who previously served as chief scientist at Salesforce. The company emerged from stealth in May with $650 million in funding at a $4.65 billion valuation.
Private, unmetered intelligence on your own machine.
Socher described the pitch for Flint in those words in the launch post. Its backers include GV, Nvidia and AMD Ventures.
