Richard Socher, now chief executive of Recursive, speaking at the Collision conference in Toronto in June 2024. Image: Piaras Ó Mídheach/Collision via Sportsfile (Collision Conf) / Wikimedia Commons, CC BY 2.0, cropped
.jpg){kind=link}
An AI startup says its own AI did most of the work on its first model, and it is giving that model away. Recursive, the company run by You.com founder Richard Socher, released Flint on Wednesday, an open-weight coding model that it says a small team of people and agents built from scratch in under four months.
The weights are free to download from Hugging Face under the Apache 2.0 licence, which allows commercial use. Recursive pitches Flint as a step towards recursive self-improvement (RSI), AI that builds better AI, an idea some lawmakers want to ban.
What Flint is #
Flint is a mixture-of-experts model made for coding agents: investigating codebases, fixing bugs, adding features and writing tests. According to its model card, it has 39.8 billion parameters in total, but only about 2.5 billion are active for each token, which is what makes it fast. It handles text only, supports reasoning and tool calls, and has a documented context window of 262,144 tokens.
Recursive’s own scores put it at 89.7% on SWE-bench Pro v2, 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1, and it claims Flint leads every mixture-of-experts model under 50 billion parameters on agentic coding. These are company-run benchmarks and haven’t been checked independently. Rivals’ figures in its charts are their own, run under different setups.
Free to download, built to run on your own machine #
The full 16-bit checkpoint is about 80 GB and runs on a single Nvidia B300 or H200 data centre GPU. For ordinary hardware, Recursive has made smaller quantised versions for Nvidia and Apple Silicon, and says Flint works with llama.cpp, Ollama, LM Studio, Unsloth, vLLM and SGLang. It has also released rec, its own lightweight engine tuned for Macs with M4 Pro and M5 Pro or Max chips.
The company claims more than 125 tokens a second on an M4 Pro MacBook Pro, up to twice as fast as similar models. Because the weights are public, Flint works anywhere in the world, offline, with code never leaving the computer, though Recursive admits it “may need more guidance” than the big closed coding models.
“Done and improved upon by AI” #
The real pitch is how Flint was made. Recursive says its automated research system, the Eureka Machine, proposes ideas, writes the code, tests them and picks what to try next, and that it helped develop “nearly every part” of Flint’s model-building stack, from data and training recipes to infrastructure and inference.
Recursive says the system picked Flint’s 2.5 billion active design, chose, against common practice, to squeeze the output layer to 4 bits, and wrote custom kernels for Apple and Nvidia chips.
Socher went further in an interview with Wired, saying “nearly every aspect of the entire model pipeline was actually done and improved upon by AI”, including a new way of training the base model. He says that is the only reason a handful of researchers could compete, a claim no one has yet checked:
It’s the only way this very small team can produce a model that’s actually better than any other model in its class.
Richard Socher, CEO and cofounder of Recursive, to Wired
A company betting on the thing Congress may ban #
Recursive is open about its goal: to “build recursively self-improving superintelligence to automate knowledge discovery”. It has raised $650 million at a $4.65 billion valuation from investors including Google Ventures, Nvidia and AMD Ventures, Wired reports.
That goal is exactly what worries safety researchers. Anthropic’s alignment lead Evan Hubinger has put the odds of AI killing all humans within a decade above 10%, with AI speeding up its own development the central fear, and even OpenAI has called for global standards before AI starts building smarter AI. The Sanders and Casar superintelligence ban backed by Alexandria Ocasio-Cortez would also outlaw recursive self-improvement, and Senator Maria Cantwell’s frontier AI framework would make companies report “unsafe recursive self-improvement”.
Recursive calls such fears more science fiction than reality, and Socher told Wired that Flint was built with a system that cannot escape human control. Even when the Eureka Machine reaches full self-improvement, he said, people will oversee it and could stop it if it goes wrong.
Why it matters #
AI building AI has mostly been debated as a future risk. Recursive says it has already shipped a model made largely that way and handed it to anyone who wants it, so the fight in Congress over banning RSI is no longer only about what might happen.
Sources: Recursive (Introducing Flint), Flint model card, Hugging Face, Wired (interview with Richard Socher).