How Recursive Says Its AI Built Flint, a Fast Coding Model for Laptops
Richard Socher's startup Recursive released Flint, an open-weight coding model it says was largely designed by its own automated research system. The self-reported benchmarks look strong, but the bigger claim is about AI building AI.
Recursive wants credit for two things at once. The first is a small coding model that runs fast on a laptop. The second is a much bigger claim: that its own AI research system, not its people, did most of the work of building it.
The startup, led by former Salesforce chief scientist Richard Socher, released Flint on Wednesday, October 21. It is the company's first model, and Recursive is pitching it as a working demonstration of recursive self-improvement, the idea that AI can build better AI. Depending on who you ask, that is either a shortcut to scientific progress or a safety nightmare.
"Nearly every aspect of the entire model pipeline was actually done and improved upon by AI." Richard Socher, CEO of Recursive, speaking to Wired
What Flint actually is
Strip away the philosophy and Flint is a mixture-of-experts model with 40 billion total parameters, of which about 2.5 billion fire for any given token. The Hugging Face model card puts the exact count at 39.8 billion. The weights are free under the Apache 2.0 license.
Speed is the headline feature. TechCrunch reports that in Recursive's tests Flint hit about 131 tokens per second on an M4 Pro MacBook Pro, roughly double the 58 to 69 tokens per second that Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning managed on the same machine.
The benchmark numbers are all Recursive's own. On its model card, the company reports Flint at 74.2% on SWE-bench Verified, 89.7% on SWE-bench Pro v2 and 54.9% on Terminal-Bench 2.1. The card itself warns that evaluation setups vary across sources. Recursive claims Flint leads mixture-of-experts models under 50 billion parameters, but by TechCrunch's reading, Alibaba's dense Qwen3.8 27B still beats it on the company's own six-test average. Nobody has independently replicated the scores yet.
The company is also candid about limits. Its launch post says Flint "may need more guidance than the larger closed-source coding models you're used to" and suggests breaking big tasks into smaller steps.
Early hands-on testing is kind. Wired's Will Knight ran Flint on a Mac with 48GB of memory and wrote that it "nailed my coding requests every time."
How much was really built by AI
This is the part worth scrutinizing. According to Recursive's announcement, its Eureka Machine proposes ideas, implements them, tests them and decides what to try next. The company credits it with work on data, optimizers, training recipes, infrastructure and inference. It picked the final 40B shape and wrote custom kernels for Apple Silicon. And on quantization, Recursive says it broke convention: "Against common practice, it quantized the output layer (lm_head) to 4 bits."
That is a meaningful list. It is also a list of contributions inside goals that people chose. Recursive frames the team as "people and agents" and has not said how much researcher time the system saved, a gap TechCrunch flagged too. "Built by AI" here means a heavily automated research loop with humans setting targets, not a model that decided to make itself.
"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, speaking to Wired
Why the framing matters
Recursive is not shy about its end goal. Its launch post says the company aims "to build recursively self-improving superintelligence to automate knowledge discovery." Wired notes that several bills in Congress propose banning the approach. Socher told Wired the system used for Flint cannot escape human control and that even full self-improvement would be overseen by people able to stop it.
The money behind the bet is serious. Recursive came out of stealth in May with $650 million at a $4.65 billion valuation, TechCrunch reports, from backers including GV, Nvidia and AMD Ventures. It previously told TechCrunch that its $410 million compute deal with Amazon Web Services was "less about headcount and more about agent count."
Flint also lands a week after Meta's far heavier Muse Spark Open, which Meta built to run on Macs and which TechCrunch says needs at least a terabyte of memory. Recursive is betting that small and machine-designed can compete. The self-improvement story is the claim that still needs outside checking.
