Recursive Superintelligence, a startup trying to build AI that improves itself, has released its first model: Flint, a free coding model small enough to run on a laptop.
The company announced Flint on Wednesday and published the weights on Hugging Face. Weights are the trained numerical settings that make up a model, so anyone can download them and run the model locally. Recursive said its own automated research system helped build nearly every part of Flint. That is the company's description of its own work, and the announcement as provided does not say which parts, or how much human oversight was involved.
The release matters for two reasons that are easy to blur together. Small coding models that run on ordinary hardware are a crowded category, and a free download is the price of entry. The more unusual claim is about process: that software did much of the research and engineering that people usually do.
That claim is also narrower than the company's name suggests. Using an automated system to help build a model is not the same as a model rewriting itself in a loop. Flint is the output of one process, not proof that the process compounds.
The source gives no benchmark scores, parameter count, or comparison with rival models. Without them, the first outside tests of Flint will decide whether this is a competitive coding tool or mainly a demonstration of how it was made.