Startup Recursive has launched Flint, a mixture-of-experts language model developed using its proprietary Eureka Machine system. The AI-driven process handled architecture design, training data, and inference tuning, completing the model in four months. Flint features 40 billion total parameters with 2.5 billion active per token.
Designed for local coding tasks, Flint supports context windows up to 512,000 tokens, with a recommended limit of 256,000. On a MacBook Pro with M4 Pro, it generates over 125 tokens per second, while an Nvidia RTX Spark with 64 GB of memory exceeds 150. The model is available on Hugging Face under the Apache 2.0 license.
The Eureka Machine optimized the model by testing various configurations, ultimately selecting the 40-billion parameter design for its performance-to-speed ratio. It also created custom compute kernels for Apple Silicon and Nvidia hardware. Recursive released its own inference engine, rec, specifically for Apple chips.
Founded by Richard Socher and Tim Rocktäschel, Recursive exited stealth in May 2026. The company previously raised $650 million at a valuation of $4.65 billion. Flint serves as the first public demonstration of the startup's automated research pipeline, which aims to eventually achieve self-improving AI.
