Recursive launches Flint, a laptop coding model it says was largely built by AI
TechCrunch Aditya Mehta
Recursive shipped Flint, a free coding model that runs on a laptop. The twist: it says AI helped build most of the model, which is the industry’s favorite scary idea.
Based on reporting by TechCrunch, Aditya Mehta — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Recursive is trying to make AI that improves itself, and its first public shot is Flint, a coding model released Wednesday. The startup says the model was built largely with help from its own automated research system, the Eureka Machine, letting a small team finish it from scratch in under four months.
That makes Flint more than another laptop-friendly coder. It is a live test of recursive self-improvement: the idea that AI can build better AI, which then builds even better AI. People in the field have been arguing about that for years, partly because the same loop that speeds progress could also get away from human control. Recursive’s version is still bounded, though. For now it is an AI research assistant under human oversight, not a machine setting off on its own.
Flint is free to download on Hugging Face under the Apache 2.0 license, and it runs on a user’s own computer instead of a cloud service. That means offline use, local code privacy, and no usage fees that tick up with every chunk of text processed. The model is a mixture-of-experts system, so it uses only 2.5 billion of its 40 billion parameters for any one task. That keeps it small enough for an ordinary laptop.
Recursive says that speed matters too. On a MacBook Pro with Apple’s M4 Pro chip, Flint generated about 131 tokens per second in the company’s tests, roughly twice the 58 to 69 tokens per second recorded by similar models on the same machine, including Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning. On Nvidia’s RTX Spark, it passed 150 tokens per second. It also did well on SWE-Bench Verified and scored behind only one model, Alibaba’s Qwen3.8 27B, across Recursive’s own average of six tests.
The more interesting part is how Recursive says Flint was made. Eureka supposedly helped with almost every step, from pre-training to post-training and inference, and the company credits it with a sparser model shape, low-bit training that cut Flint by 71% with almost no loss in quality, and custom kernels that speed up computation. Recursive has not said how much time that saved. But the direction is clear: if the loop works, each model should help build the next one faster than the last.
My take — AI-written commentary, not fact-checked reporting
This is the cleanest version of the AI hype machine: build tools that help build the next tools, then call it research. Recursive is betting that automation will scale faster than caution, which is very Silicon Valley and only slightly less optimistic than assembling a jet engine in a garage. The real story is not that AI can code on a laptop; it is that everyone now wants AI to do the lab work too.
Read more about this at: TechCrunch