Recursive launches Flint, an open coding model that runs locally

A coding model that runs on your machine

Recursive has released Flint, a coding model designed to run on local hardware. According to the company, it was built from scratch in under four months by a small team of humans and AI agents.

Architecture and speed

Flint has 40 billion parameters in total. Thanks to its mixture-of-experts (MoE) architecture, only 2.5 billion parameters are activated per token, which reduces the compute load. In Recursive's own tests, the model generates more than 125 tokens per second on an M4 Pro MacBook Pro with short contexts.

Built with the Eureka Machine

The model was developed using the company's automated research system, Eureka Machine. The system automates idea generation, running experiments and planning new attempts based on results, and took part in stages ranging from architecture choices to training methods and speed improvements.

Open weights and limits

Flint's weights are shared on Hugging Face under the Apache 2.0 license. Local use enables offline operation and keeps data on the device while eliminating API fees. Recursive recommends breaking complex tasks into small steps and acknowledges that Flint may need more steering than larger closed models. The company says Flint beats MoE models with fewer than 50 billion parameters on certain coding tests; these results are based on the company's own evaluations.

The bigger picture

Led by Richard Socher, Recursive raised $650 million at a $4.65 billion valuation in May from investors including GV, Nvidia and AMD Ventures. The company positions Flint as the first step in a research loop in which each new model contributes to building the next one.