Recursive launches Flint coding model developed with its own AI system

Recursive Superintelligence has launched Flint, its first coding model, saying an automated research system contributed to almost every aspect of its development. The startup says Flint can run locally on a laptop and was built from scratch by a small group of humans and AI agents in less than four months.

The model’s weights were published on Hugging Face with an Apache 2.0 license, permitting free use and modification. According to Recursive, its Eureka Machine generated and tested ideas, then used the findings to guide further work on training data, infrastructure, model design and execution code.

Flint uses a mixture-of-experts architecture: it contains 40 billion parameters but activates only 2.5 billion for each word or code fragment. Recursive says the development system also compressed the output layer to 4 bits and produced specialized code for Apple and Nvidia chips.

The company reports output speeds above 125 tokens per second on a MacBook Pro equipped with an M4 Pro chip, and above 150 on Nvidia RTX Spark laptops. Separately, Nvidia said the custom code improved Flint’s speed by 55% on its RTX Spark machines without altering the model.

Recursive reported scores of 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1 in its own testing. Those results have not been independently verified by TNW, and the model card cautions that testing configurations differ. Recursive also warns that Flint can generate incorrect or harmful material and defective code, and may require more guidance than larger closed coding models.