Recursive releases 40B open-weight model for offline local coding

Recursive released Flint, its first open-weight coding model, designed to run locally so developers can work offline, keep code on-device and avoid cloud API fees. The 40-billion-parameter mixture-of-experts model activates 2.5 billion parameters per token; Recursive says it can generate more than 125 tokens per second on an M4 Pro MacBook Pro and performs strongly against other smaller coding models on its benchmarks. Built in under four months with substantial help from Recursive’s human-supervised Eureka Machine, Flint is available under the Apache 2.0 license. The company cautions that the model may need more explicit guidance than larger commercial coding assistants, while presenting the release as an early, limited example of AI helping develop AI.