By Hamza Faisal
AI startup Recursive has released Flint, a free coding model that runs on an ordinary laptop — and, according to the company, was largely built by another AI system. The release, announced on October 21, is the startup's first model and its first public test of an idea the industry calls recursive self-improvement: AI that helps build better AI.
Recursive says its automated research system, the Eureka Machine, contributed to nearly every stage of Flint's development, from searching for the right model architecture to tuning low-level GPU code. A small human team oversaw the work, and the model went from scratch to release in under four months.
"Nearly every aspect of the entire model pipeline was actually done and improved upon by AI," chief executive Richard Socher said at launch. Socher previously served as chief scientist at Salesforce, and Recursive emerged from stealth in May with $650 million in funding at a $4.65 billion valuation, backed by investors including GV, Nvidia and AMD Ventures.
Small enough for a laptop, fast enough to matter
Flint is a mixture-of-experts model, which means it activates only a slice of itself for each task: about 2.5 billion of its roughly 40 billion parameters per token. That design keeps it light enough for consumer hardware. In Recursive's tests, Flint generated about 131 tokens per second on a MacBook Pro with Apple's M4 Pro chip — roughly twice the pace of comparable models such as Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning — and topped 150 tokens per second on Nvidia's RTX Spark platform.
Because Flint runs locally, developers can work offline, keep their code on their own machines and avoid cloud usage fees. It is free to download on Hugging Face under the Apache 2.0 open-source licence, and works with local tools including Ollama, LM Studio, llama.cpp, vLLM and SGLang. Recursive is also releasing a lightweight inference engine, called rec, for Apple's M4 Pro and M5 Pro and Max chips.
On SWE-Bench Verified, a benchmark that measures how well models fix real-world software bugs, Recursive says Flint outperformed other models with fewer than 50 billion parameters. The company is careful to note a caveat of its own: Flint may need more guidance than the larger closed-source coding models developers are used to.
An old idea, now a public experiment
Recursive self-improvement has been debated for years. In its strongest form, it imagines AI systems improving themselves in a loop beyond human oversight — an idea some researchers find exciting and others alarming. Recursive's version is far more modest: an AI research assistant, overseen by people, that produced a laptop-sized model.
Flint is not a frontier model competing with the largest systems from OpenAI, Anthropic or Google. It is a demonstration that an automated research pipeline can carry a real model from architecture search to release — and that the result can be genuinely useful, fast and free.
What it means for developers
The pitch, in Socher's words, is "private, unmetered intelligence on your own machine." If tools like Flint keep improving, the economics of coding assistance could shift: developers would no longer need to send every keystroke to a cloud provider, pay per token, or worry about proprietary code leaving their laptop.
Whether automated research pipelines can scale from a laptop model to frontier-class systems remains an open question, and Recursive's speed claims are its own test results, not independent measurements. But Flint is now public, open and free — anyone with a laptop can download it and find out.
