Recursive Debuts Flint, an AI-Built Coding Model for Laptops

Flint functions as an early, controlled test of recursive self-improvement—the concept that AI can iteratively build more capable successors. Unlike cloud-based assistants that require subscription fees and constant internet connectivity, Flint runs locally on a user’s hardware. This architecture ensures code privacy and offline functionality, making it a distinct alternative to larger, server-dependent models. The model is available on Hugging Face under an Apache 2.0 license.

Technical efficiency defines the model’s performance. As a mixture-of-experts system, it activates only 2.5 billion of its 40 billion parameters for any given task. This compact design allows it to run on standard laptops, achieving speeds of 131 tokens per second on a MacBook Pro with an M4 Pro chip. In performance benchmarks, such as SWE-Bench Verified, Flint outperformed competitors within its size class, proving effective for tasks ranging from building simple applications to complex bug fixes.

The Eureka Machine contributed to the model’s architecture, including low-bit training that reduced its size by 71% without sacrificing quality. While the company admits that users may need to break down large projects into smaller steps, early testing indicates the model handles coding requests with high accuracy. Recursive plans to continue this trajectory, aiming for a future where each successive model enhances the Eureka Machine’s research capabilities, further shortening the development cycle for future iterations.