AI Assistant Builds Flint Model

An automated research assistant called the Eureka Machine has helped build a new open weight coding model named Flint in under four months. By proposing ideas, running experiments, and refining software architectures under human oversight, the platform enabled a small team to complete the build from scratch. The system contributed across development stages, while low-bit quantization and custom computation kernels reduced the model’s memory footprint by 71 percent with minimal performance loss. This approach showcases how supervised automated research can accelerate the development of specialized software tools.

Flint employs a mixture-of-experts architecture that activates about 2.5 billion of its 40 billion total parameters per token. During evaluations on a 64GB Nvidia RTX Spark GPU at short context lengths, the model generated more than 150 tokens per second. On an Apple M4 Pro chip, it reached more than 125 tokens per second in Recursive’s internal tests. On the SWE-Bench Verified bug fixing evaluation, Recursive reports that Flint led all mixture-of-experts models under 50 billion parameters.

To make the tool widely available, Recursive published the model weights for free on Hugging Face under the Apache 2.0 license. Operating locally on compatible laptops helps keep code and developer inputs on the user device without incurring usage-based cloud API charges. This offline capability offers engineers an efficient local option for handling software development tasks.