Recursive Unveils Flint, a Laptop-Ready Coding AI Built Largely by AI Itself

Image: TechCrunch · Source

AI startup Recursive launched Flint, a coding model designed to run on laptops and created mainly using its AI research system. Flint exemplifies early recursive self-improvement, showcasing AI-assisted AI development with promising performance and accessibility.

San Francisco-based AI startup Recursive has introduced Flint, a novel coding AI model that represents an innovative step in recursive self-improvement — the concept of AI systems creating improved AI models autonomously. Launched on October 21, 2026, Flint is a free, lightweight coding model designed to run efficiently on ordinary laptops, marking a notable achievement in AI development where a small human team leveraged Recursive’s AI research engine, dubbed the Eureka Machine, to create it in under four months.

The Eureka Machine is an automated research system that contributed to nearly every stage of Flint’s creation, from pretraining through to optimization and inference. Recursive attributes key advancements—such as a sparser model architecture, a 71% size reduction via low-bit training with minimal quality loss, and the development of specialized computational kernels—to this AI-driven process. By integrating AI in its own research workflow, Recursive illustrates a practical example of recursive AI improvement while maintaining human oversight for safety and control.

Flint features a mixture-of-experts architecture that activates only about 2.5 billion out of its total 40 billion parameters for any given task, enabling it to run with modest computational resources. This design contrasts sharply with much larger models like Meta’s Muse Spark Open, which requires at least a terabyte of memory for its 816 billion parameters. Benchmarks demonstrate Flint’s efficiency and speed, producing around 131 tokens per second on a MacBook Pro with Apple’s M4 Pro chip, outperforming comparable models such as Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning.

Performance-wise, Flint has shown competitive results on the SWE-Bench Verified benchmark for fixing real-world software bugs, outperforming other models under 50 billion parameters and ranking just below Alibaba’s Qwen3.8 27B in Recursive’s tests. Despite requiring more user guidance for complex tasks than larger commercial coding assistants, Flint’s accuracy and responsiveness have earned positive early reviews. For instance, Wired reporter Will Knight successfully used Flint to develop small applications, noting its reliability and speed compared to smaller coding models.

By enabling developers to run Flint on local hardware, Recursive offers benefits including offline operation, enhanced code privacy, and the avoidance of cloud-based usage fees. The model is openly accessible through Hugging Face under the Apache 2.0 license, supporting free use and modification.

Recursive’s ambitious vision encompasses further advancing recursive AI systems to autonomously carry out longer research cycles with decreasing human intervention, while ensuring humans can monitor and interrupt the process as needed. The company’s substantial funding—$650 million raised since its stealth inception, backed by prominent investors like GV, Nvidia, and AMD Ventures—and strategic partnerships including a $410 million compute agreement with Amazon Web Services underscore its commitment to this approach.

With Flint as a tangible demonstration of AI-assisted AI development, Recursive aims to refine its Eureka Machine into a more capable researcher that accelerates the creation of future, more sophisticated AI models. This progress marks a significant milestone in the ongoing exploration of recursive self-improvement within artificial intelligence.