Recursive releases Flint, an AI-driven MoE model for agentic coding on local hardware

AI startup Recursive has released Flint, its first language model, which was developed using an AI system called "Eureka Machine" that autonomously proposed, ran, and evaluated experiments. Flint is a Mixture-of-Experts (MoE) model with 40 billion total parameters but only 2.5 billion active parameters per token, designed for agentic coding on local hardware. Recursive claims Flint leads coding benchmarks like SWE-Bench Pro v2 and Terminal-Bench 2.1 among MoE models with fewer than 50 billion parameters, matching the benchmark scores of much slower dense models in the 27 to 31 billion parameter range on an M4 Pro with 48 GB of memory. On a MacBook Pro with M4 Pro, it generates over 125 tokens per second at short contexts, and over 150 on an Nvidia RTX Spark with 64 GB of memory.