Recursive, an artificial intelligence startup founded in 2023, introduced Flint, a free coding model designed to operate on standard laptops. The company described Flint as an early demonstration of recursive self-improvement, where AI systems help develop more capable AI tools. Unlike many AI models that require cloud infrastructure, Flint runs locally, allowing developers to work offline and maintain code privacy without incurring API fees.

The model was developed using Recursive’s automated research system, internally referred to as the Eureka Machine. According to the company, this system contributed to nearly all stages of Flint’s creation, enabling a small team to complete the project in under four months. Flint contains 40 billion parameters but activates only 2.5 billion during any given task, a design choice that reduces computational demands and makes it suitable for consumer-grade hardware.

In benchmark testing, Flint processed approximately 131 tokens per second on a MacBook Pro equipped with Apple’s M4 Pro chip, outperforming comparable models such as Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning, which averaged between 58 and 69 tokens per second on the same hardware. On Nvidia’s RTX Spark platform, Flint achieved over 150 tokens per second. It also ranked among top-performing models under 50 billion parameters on SWE-Bench Verified, a test measuring software bug resolution capabilities, trailing only Alibaba’s Qwen3.8 27B in Recursive’s internal evaluations.

Flint is distributed under the Apache 2.0 license via Hugging Face, permitting free use and modification. Its efficiency and open availability distinguish it from larger models like Meta’s Muse Spark Open, which requires at least a terabyte of memory and is not designed for typical laptops. Early user testing reported by Wired’s Will Knight demonstrated Flint’s ability to generate functional code for tasks such as building a to-do list application and a web research tool, though it occasionally required task segmentation for complex projects.

Recursive’s research process relies heavily on AI-driven iteration, with the Eureka Machine proposing structural changes, optimizing training methods, and generating specialized code to accelerate computations. The company noted that such systems are increasingly integral to AI development, citing internal data indicating that AI agents had logged 3.1 workdays for every one worked by human researchers by mid-August of the previous year.

Looking ahead, Recursive aims to refine Flint through iterative cycles of AI-assisted development, with the long-term objective of achieving full recursive self-improvement under human oversight. CEO Richard Socher, formerly chief scientist at Salesforce, emphasized that future models should operate with increasing autonomy while remaining accountable to human control mechanisms. The company has not disclosed the extent of time saved by researchers through AI collaboration, but affirmed that its $410 million agreement with Amazon Web Services supports not headcount growth but expanded agent-based research capacity.

Flint’s release marks a notable development in accessible AI development tools, particularly for developers prioritizing privacy, offline functionality, and cost efficiency. Its performance relative to larger, resource-intensive models continues to fuel discussion around the scalability and control of self-improving AI systems.