Startup Recursive has released Flint, a 40-billion-parameter open-weight coding model that it says was largely developed by another AI system. Its internal “Eureka Machine” helped shape training methods, architecture and code optimized for particular chips.
Recursive says a small research team built the model in a few months and that Flint beats similarly sized models on coding benchmarks. A WIRED reviewer ran it locally on a Mac with 48 GB of memory and reported strong results on several small programming tasks. Those hands-on examples are encouraging but are not a comprehensive independent evaluation.
The company’s larger goal is recursive self-improvement: using AI to choose and implement changes to the systems that build later AI models. Chief executive Richard Socher says humans can oversee and stop the process, while critics argue that increasingly autonomous improvement could become hard to control.
Recursive has raised $650 million at a reported $4.65 billion valuation. Flint shows that automation can help a small team build a competitive model, but the company has not established that improvement can continue indefinitely or safely without substantial human judgment.