Recursive Releases Open-Weight Flint Model Built by AI

AI startup Recursive has launched Flint, a 40-billion-parameter open-weight coding model created through its recursive self-improvement framework, the Eureka Machine. The company, co-founded by former Salesforce chief scientist Richard Socher and University of British Columbia professor Jeff Clune, recently raised $650 million at a $4.65 billion valuation backed by Google Ventures, Greycroft, Nvidia, and AMD Ventures.

According to Socher, the Eureka Machine handled nearly the entire development pipeline for Flint, creating new training methods, architectural designs, and chip-specific optimizations. Benchmark results shared by the startup indicate that Flint surpasses competing models of similar scale on programming tasks. By leveraging automated feedback loops, the small team managed to build a competitive model in just a few months.

For practitioners, Flint brings reliable local coding performance to modest hardware. When evaluated on a Mac equipped with 48 GB of RAM, the 40-billion-parameter model generated functional web scrapers and command-line tools cleanly, avoiding the speed and error pitfalls typical of mid-sized open-weight models compared to commercial services like Claude or Codex. While trillion-parameter industrial models require server racks, Flint shows that self-improving pipelines can yield highly efficient local alternatives.

The development arrives during broader discussions regarding recursive systems. Safety debate flared in September when an Anthropic researcher quit over extinction concerns linked to self-improving models, driving proposals in Congress to enact bans. Meanwhile, research from Microsoft, Carnegie Mellon University, and the University of Virginia demonstrates that recursive self-modification is technically feasible, pointing toward a shift in how software engineers and researchers train specialized models.