Recursive, led by ex-Salesforce chief scientist Richard Socher, released Flint—a free local coding model largely developed via the Eureka Machine automated research loop in under four months.
Flint is a 40B-parameter mixture-of-experts model activating 2.5B parameters per task, downloadable under Apache 2.0. Recursive reports ~131 tokens/sec on an M4 Pro MacBook Pro and >150 tokens/sec on Nvidia RTX Spark, beating Gemma 4 and Nemotron 3.5 Lightning in its tests. SWE-Bench Verified leads sub-50B peers.
Eureka contributed sparser architecture, 71% size reduction via low-bit training, and custom kernels. CEO Socher told Wired the long-term goal is supervised recursive self-improvement. The company raised 4.65B valuation and signed a $410M AWS compute deal emphasizing agent count over headcount.