Recursive releases Flint for local agentic coding

What is changing?

Recursive has released Flint, a model built for agentic coding on local hardware. According to the company’s model card, the decoder-only mixture-of-experts architecture has 39.8 billion total parameters, with roughly 2.5 billion active for each token. It is designed to investigate codebases, fix bugs, implement features, and write tests.

Flint’s BF16 weights are available on Hugging Face under the Apache 2.0 license. Recursive says separately quantized variants are being prepared for Apple Silicon and NVIDIA hardware, and that the model supports multi-turn tool use. The company reports context windows up to 512,000 tokens, while recommending users stay within 256,000 tokens for the best results.

Recursive says its automated research system, Eureka Machine, contributed across much of the development pipeline, including data, training recipes, architecture, infrastructure, and inference optimization. The approach could expand the role of automation in helping small teams build specialized local models. However, the published benchmarks rely on developer measurements, and the model card warns that faulty code and unsafe tool calls remain possible, recommending testing and strict permission boundaries.