Artificial intelligence startup Recursive has introduced Flint, a free open-weight coding model designed to run locally on compatible laptops. Available on Hugging Face under the Apache 2.0 license, the software allows developers to work offline while avoiding cloud API fees. Running the software locally helps keep code and project inputs on the user device instead of transmitting data to external servers. Developed from scratch in under four months by a small team, the project represents a practical step toward human-supervised research automation.
The system uses a mixture of experts architecture that activates approximately 2.5 billion of its 40 billion total parameters per token. This design reduces the amount of memory that must be actively used during inference compared with a dense model of similar total size. In internal tests conducted by Recursive on a MacBook Pro powered by an M4 Pro chip, the model generated approximately 131 tokens per second.
Development of the model relied heavily on an automated research system known as the Eureka Machine. Working as an assistant under human oversight, this system contributed across pretraining, posttraining, and inference optimization phases. Through low bit quantization and custom computation kernels, the team reduced the memory footprint by 71 percent with minimal loss in benchmark performance. According to Recursive evaluations, Flint led mixture of experts models under 50 billion parameters on the SWE-Bench Verified coding benchmark.