Recursive released Flint, a 40-billion-parameter model for local coding, and says its Eureka Machine contributed across the model-development process.
Recursive released Flint on October 21, its first model built for coding on local hardware. The startup describes Flint as a mixture-of-experts model with 40 billion total parameters, activating 2.5 billion for each token. Its weights are available under the Apache 2.0 license, enabling developers to download and adapt the model.
The release also serves as a public demonstration of Recursive’s Eureka Machine, an automated research system. Recursive says the system helped with nearly every part of Flint’s development, including data, training recipes, infrastructure and inference optimization, and that a small team built the model from scratch in under four months. Those accounts and performance claims come from the company; they do not establish that Flint was developed autonomously.
For local use, Recursive reports generation speeds above 125 tokens per second on an M4 Pro MacBook Pro and above 150 tokens per second on a 64GB NVIDIA RTX Spark at short contexts. The company says Flint leads mixture-of-experts models below 50 billion parameters on several coding benchmarks, including SWE-Bench Verified, and can rival similarly sized dense models while running more than five times faster. These are vendor-reported results, so comparisons may depend on evaluation setup and hardware.
Running a coding model on a developer’s own machine can reduce reliance on cloud services, keep code on-device and avoid usage-based API charges. Recursive also cautions that Flint may need more guidance than larger closed-source coding assistants, particularly for complex work. The launch therefore combines a practical local-model release with an early test of AI-assisted model research, while leaving broader claims about self-improving systems unproven.