An AI programmer you can run right on your laptop

Recursive has unveiled Flint, a free code generation model that can be used locally on laptops without a cloud connection. The development took less than four months, with the Eureka Machine automated research system playing a significant role, while the entire process was overseen by human experts throughout.

Photo: Pexels

Startup company Recursive has released a model on the Hugging Face platform under the Apache 2.0 license FlintIt is a demonstration of how advanced systems can help create new models of artificial intelligence. The Eureka Machine system suggested experiments and guided development, but only in the role of a research assistant under the supervision of a team of experts.

The main advantage of local operation is that the program code and data remain on the user's computer, without being sent to the cloud and without paying subscription fees based on the amount of text processed. The disadvantage is the need for appropriate hardware and the fact that the model requires much more detailed instructions than established paid cloud assistants.

Flint uses a “mixture-of-experts” architecture with a total of 40 billion parameters, with only 2.5 billion activated for each task. In internal tests on a MacBook Pro with an Apple M4 Pro chip, it achieved a speed of around 131 tokens per second, which is noticeably faster than models such as Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning.

On the Nvidia RTX Spark graphics card, the speed exceeded 150 chips per second. According to the company, in the SWE-Bench Verified test, it surpassed the results of competing models by less than 50 billion parameters. Despite the promising numbers, a certain amount of caution should be maintained, as the detailed testing conditions were not publicly disclosed, so the actual usefulness of the model will only be confirmed by wider use in practice.

Recursive sees the new Flint as just a first step. While the vision of automated AI development is appealing, Flint is primarily a demonstration that it is possible to create a powerful and manageable local model for generating code.