Recursive: flint-v1.0:free

Flint is Recursive's first model: an open-weight mixture-of-experts language model built for coding agents. It has 39.8 billion total parameters with approximately 2.5 billion active per token, arranged across 49 layers — 2 dense layers followed by 47 mixture-of-experts layers. It accepts text input and produces text output, supports multi-turn tool use, and ships under the Apache 2.0 license, so it can be used commercially, modified, and redistributed. Recursive says Flint was largely built by its own AI research system, Eureka Machine. The model is aimed at agentic coding workflows: investigating codebases, fixing bugs, implementing features, and writing tests. Its 262,000-token context is shared between prompt and output, enough to hold a meaningful slice of a mid-sized repository alongside the agent's reasoning trace without aggressive chunking or retrieval tricks; Recursive recommends an output budget of at least 32,768 tokens. Typical uses include helping developers write code, building small applications, and solving programming problems. Flint can run on a single B300 or H200 GPU per replica using an SGLang serving kit, and Recursive reports throughput of 125 or more tokens per second on an M4 Pro and 150 or more on a 64GB NVIDIA GPU. The BF16 checkpoint weighs around 80 GB. It was free to access on OpenRouter at launch.