The model matters less than the machine that made it. Recursive, a startup founded last year to build AI that improves itself, released Flint on Wednesday, a free coding model that runs on a laptop. The company says its automated research system, the Eureka Machine, helped with nearly every step of building it, which let a small team finish from scratch in under four months.
Flint is a mixture-of-experts model with 40 billion parameters, but it activates only 2.5 billion at a time. That is why it fits on an ordinary laptop. Compare that with Meta’s Muse Spark Open, released last week for Macs, which uses 816 billion parameters and needs at least a terabyte of memory. Those are very different bets on what local AI should look like.
Speed is the other selling point. In Recursive’s own tests, Flint produced about 131 tokens per second on a MacBook Pro with an M4 Pro chip. Similar models, including Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning, managed 58 to 69 on the same machine. On Nvidia’s RTX Spark it passed 150. These are vendor numbers, so treat them with some caution, but the gap is large.
On quality, Flint beat other sub-50-billion-parameter models on SWE-Bench Verified, which tests real-world bug fixing. Across Recursive’s six-test average, only Alibaba’s Qwen3.8 27B scored higher. The company admits Flint needs more hand-holding than big paid assistants and recommends splitting large jobs into smaller steps. Still, Wired’s Will Knight ran it on a Mac with 48GB of memory and said it handled every request, including a to-do app and a web research tool. Smaller models, he noted, usually stumble or crawl compared with Claude or Codex.
The practical case is easy to follow. Flint runs offline, keeps code on your machine, and carries no per-token fees. It is free on Hugging Face under Apache 2.0, so you can modify it and use it commercially. For teams wary of sending proprietary code to a cloud API, that is a real option, though it won’t replace frontier coding agents for hard work.
The Eureka Machine is the actual story. It proposes ideas, runs experiments, and uses the results to pick what to try next. Recursive says it contributed to pretraining, post-training, and inference. Its credited wins include a sparser model shape, low-bit training that shrank Flint by 71% with almost no quality loss, and custom kernels that speed up computation. Recursive didn’t say how much researcher time this saved, which is the number I would most like to see.
Recursive isn’t alone in automating research. OpenAI said last month that by mid-August its AI agents were logging 3.1 workdays for every one worked by a human researcher. The idea of AI building better AI is one of the most contested in the industry, with some researchers fearing it could slip out of human control. Recursive’s version is modest for now: a human-supervised research assistant that produced a small model.
The backing is not modest. Recursive came out of stealth in May with $650 million at a $4.65 billion valuation, from investors including GV, Nvidia, and AMD Ventures. CEO Richard Socher, Salesforce’s former chief scientist, told Wired the long-term goal is full recursive self-improvement, with humans overseeing it and able to stop it. The company also signed a $410 million compute deal with AWS in June and described it as “less about headcount and more about agent count.”
So is this noise? The self-improvement pitch is still mostly a claim. But Flint is a usable, open, fast local model, and that part is real. Download it and test it on your own codebase before believing any benchmark, mine included.
