Recursive releases Flint 40B local coding model built using automated AI

Artificial intelligence startup Recursive released its first model, Flint, on October 21, 2026. TechCrunch AI and Wired AI both report that the free coding model was built largely using the company's automated research system, known as the Eureka Machine. MarkTechPost notes that the system is an open-weight 40B mixture-of-experts model with 2.5B active parameters released under an Apache 2.0 license. [1][2][3]

Recursive claims the model can run locally on laptops. MarkTechPost reports that Flint achieves 125 or more tokens per second on an M4 Pro MacBook Pro. Outlets differ on benchmark performance, with MarkTechPost citing 74.2% on SWE-bench Verified, while Tech Insider reports an 89.7% score on SWE-Bench. [1][5]

According to Wired AI, Recursive chief executive officer and co-founder Richard Socher stated that AI developed and improved nearly every step of the model pipeline. Wired AI and TechCrunch AI report that the automated process allowed a small team of researchers to construct the model in a few months. [2][3]

FourWeekMBA reports that Recursive has raised 650 million dollars in backing. The company aims to use recursive self-improvement methods to advance both artificial intelligence systems and scientific research. [3][6]