Recursive launches Flint, an open 40B coding model it says AI largely built

Startup Recursive released Flint, a free coding model meant to run on a laptop rather than in the cloud. It is a mixture-of-experts model with 40 billion total parameters but only activates 2.5 billion of them per task, which is why it can run on ordinary hardware. The weights are posted on Hugging Face under the Apache 2.0 license, so anyone can download, run offline, and modify it.

Recursive says the model was built in under four months by a small team aided heavily by its automated research system, called the Eureka Machine, which proposed and tested ideas across data, training, and the inference stack. The company reports (its own benchmarks) that Flint leads other mixture-of-experts models under 50 billion parameters on coding tests such as SWE-Bench Verified, and that it runs over 125 tokens per second on an Apple M4 Pro laptop and over 150 tokens per second on an Nvidia RTX Spark machine, faster than similarly sized models according to Recursive's testing.

TechCrunch reported that Wired's reviewer found Flint handled simple coding tasks reliably on a Mac, though Recursive itself says the model needs more hand-holding than larger paid coding assistants. Recursive previously raised $650 million and is led by former Salesforce chief scientist Richard Socher.

Why it matters

It's a free, open, laptop-capable coding model built largely by AI research tools, a real-world test of AI systems improving AI development.

Source: Recursive