What is Flint
According to a report by TechCrunch, the startup Recursive published Flint on Wednesday, a free coding model that can run on a laptop. The company says the model was built largely with the help of its own artificial intelligence, the automated research system it calls Eureka Machine, and that a small team managed to build it from scratch in less than four months. Flint is available for download on Hugging Face under the Apache 2.0 license, which allows free use and modification.
Its technical architecture is mixture-of-experts: for each task, the model activates only a fraction of its 40 billion parameters, amounting to 2.5 billion. This choice makes it light enough for a laptop and much smaller, again according to the comparison offered by the source, than Meta's Muse Spark Open, which uses 816 billion parameters and requires at least a terabyte of memory.
Claimed performance and comparison
Recursive claims that Flint generates about 131 tokens per second on a MacBook Pro with an Apple M4 Pro chip, roughly double the 58-69 tokens per second attributed to similar models on the same machine, including Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning; on an Nvidia RTX Spark it would exceed 150 tokens per second. These figures come from the company's tests and not from independent verification.
According to the company, Flint would outperform other models under 50 billion parameters on the SWE-Bench Verified benchmark, which measures the ability to fix real bugs in software. Averaged across six internal Recursive tests, only Alibaba's Qwen3.8 27B is said to have scored higher. The company also admits that Flint requires more guidance than large paid assistants and suggests breaking complex tasks into smaller steps. The source also cites a positive trial by Wired journalist Will Knight, who reportedly used Flint on a Mac with 48 GB of memory to build small programs.
Why it matters and what remains to be verified
The point of interest is not only the model but the method: Flint is presented as an initial public test of so-called recursive self-improvement, the idea that AI can build better AI. Recursive says Eureka Machine contributed to almost all stages, from pre-training to post-training through to inference, suggesting a more sparse model architecture, low-precision training that reportedly shrank Flint by 71 percent with almost no quality loss, and custom kernels. The company did not say how much time the system saved researchers.
The industry debate over this approach is open: the source notes that some researchers fear an AI capable of improving itself could escape human control, while Recursive's current version remains a research assistant supervised by humans, with the ability to stop the process. The stated goal of CEO Richard Socher is to reach full recursive self-improvement, but this is an announced and unproven prospect. It therefore remains to independently verify the claimed performance, Eureka Machine's actual contribution, and how well the model holds up on coding tasks more complex than those cited.
Who is behind it
Recursive emerged from stealth in May with $650 million in funding and a $4.65 billion valuation, with backers including GV, Nvidia and AMD Ventures. The company is led by Richard Socher, former chief scientist at Salesforce. The source also notes a $410 million compute agreement with Amazon Web Services signed in June, which the company had described as focused more on the number of agents than on the number of employees.
Flint's release should be read in this context: a free, local model allows developers to work offline, keep their code private, and avoid the usage costs of cloud services calculated based on the text processed. These are advantages described by the source as product features, not as results measured by third parties.
