Recursive releases Flint, an open coding AI model it says AI largely built

Recursive, a US artificial intelligence start-up, released its first AI model, called Flint, on Wednesday, October 21, 2026. Flint is a coding model, a program that writes and fixes software code, and it is small enough to run on a laptop without an internet connection. The company made its weights, the trained numbers that make up the model, free to download on Hugging Face, and it says an automated research system of its own, called the Eureka Machine, carried out most of the work of building it.

Most of the coding assistants that software developers use today, in India and elsewhere, run on large models in distant data centres. Each request is sent over the internet, and companies pay per use, usually counted in tokens, the small chunks of text a model reads and writes. A model that runs on a developer's own laptop changes that arrangement: the code never leaves the machine, there is no bill per request, and it works offline. That matters to companies that handle sensitive code, such as banks and government contractors, and to developers in places where fast internet and foreign-currency subscriptions are hard to come by.

The bigger claim is about how Flint was made. Recursive says its automated research system did much of the work human AI researchers normally do: proposing training methods, trying model designs and writing code optimised for specific chips. Richard Socher of Recursive described the Eureka Machine as having come up with several new ideas in the process. If such systems really can speed up AI research, the pace of new models could rise sharply. But this is the company's own account. Independent reviewers who tried the model reported that it handled basic coding requests well, while noting that the benchmark results were supplied by the company and have not yet been checked by others.

A mixture of experts is a design in which a model is split into many smaller expert networks and a router picks only a handful for each piece of text. The model can hold a lot of knowledge while using only a small part of its computing power at any moment. Open weights mean anyone can download, run and adapt the model, unlike closed models that can only be used through a company's paid service.

Short-term: Developers can download Flint now and test it against the cloud assistants they already pay for. Indian IT services firms and start-ups that cannot send client code to outside servers have one more option for an offline coding helper, though they will need to run their own tests before relying on it.

Long-term: Flint adds to a fast-growing group of open models that run on personal hardware, a trend that lowers the cost of using AI and reduces dependence on a few large providers. If Recursive's claim about automated research holds up under outside scrutiny, it would strengthen the argument that AI systems can speed up the building of other AI systems, a development regulators and safety researchers are watching closely.

Who is affected: Software developers and students learning to code, companies with strict data rules, makers of paid coding assistants who face cheaper competition, and laptop and chip makers whose hardware the model is tuned for.

Questions and Answers