Recursive Unveils Flint: A Laptop‑Ready Coding AI Mostly Built by Machines

When you hear a startup brag about a model that “writes code on your laptop,” you might expect a thinly‑veiled marketing gimmick. Instead, Recursive, the AI‑first company that popped out of stealth in May, actually delivered something you can download, run offline, and—yes—tinker with.

On Wednesday the team dropped Flint, a free‑to‑use coding model that fits comfortably on a typical notebook. What makes Flint oddball isn’t just its size; it’s the way it got built. Recursive says a self‑directed research system they call the “Eureka Machine” handled almost every step of development, from sketching the architecture to polishing the final inference code. In plain English, a bunch of algorithms did most of the heavy lifting, letting a handful of engineers finish the job in under four months.

Flint is what AI researchers call a mixture‑of‑experts model. Instead of waking up all 40 billion of its parameters for each request, it awakens just a 2.5‑billion‑parameter slice. That’s enough to be useful, yet light enough to run on an Apple M4 Pro or even a modest Windows laptop without choking the system. By contrast, Meta’s newly‑released Muse Spark Open carries a whopping 816 billion parameters and needs a terabyte of RAM—hardly a “laptop‑friendly” proposition.

Speed matters, too. In Recursive’s own tests Flint cranked out roughly 131 tokens per second on a MacBook Pro with the M4 Pro chip—about twice the pace of comparable models like Google’s Gemma 4 or Nvidia’s Nemotron 3.5 Lightning on the same hardware. On an Nvidia RTX Spark GPU the model topped 150 tokens per second, which is impressive for something that fits in a laptop’s memory.

Performance isn’t just about raw throughput. On the SWE‑Bench Verified benchmark—an industry‑standard test for fixing real‑world code bugs—Flint beat every other model under 50 billion parameters. Only Alibaba’s Qwen‑3.8 27B nudged it out of the top spot in Recursive’s internal ranking across six different tests.

Early users seem to agree. Wired’s Will Knight gave Flint a whirl on a Mac with 48 GB of RAM, asking it to spin up a tiny to‑do list app and a web‑scraping utility. He notes that smaller coding models often stumble or lag behind the big‑name assistants like Claude or Codex, but Flint “nailed” his prompts each time. The caveat? Flint still needs more human guidance than the polished, paid‑for assistants; breaking big tasks into bite‑size steps helps it stay on track.

How did Recursive manage to squeeze all this out of a modest research budget? The answer circles back to the Eureka Machine. According to the launch video, the system generated ideas, ran experiments, and iterated on results throughout the entire pipeline—pre‑training, fine‑tuning, even the custom kernels that accelerate the math under the hood. One concrete win was a low‑bit training regime that shrank the model by 71 percent with barely any quality loss.

The company’s backers are optimistic, too. Recursive emerged from stealth with a $650 million Series B round led by GV, Nvidia, and AMD Ventures, valuing the startup at $4.65 billion. CEO Richard Socher—formerly Salesforce’s chief scientist—tells Wired the long‑term vision is “full recursive self‑improvement,” where each new model makes the Eureka Machine a better researcher, eventually allowing the system to drive its own evolution with humans only stepping in to pull the emergency brake.

For now, Flint lives on Hugging Face under an Apache 2.0 license, meaning anyone can download, modify, or ship it without paying royalties. That openness, combined with the ability to run offline, sidesteps the usual cloud‑AI fees that charge by the token processed. In other words, developers can keep their code private, avoid costly usage bills, and still get a decent AI coding buddy.

Whether Flint is a glimpse of a future where AI continually improves itself—or just a clever proof‑of‑concept—remains to be seen. What’s clear is that Recursive has managed to turn a lofty idea into something you can actually run on your own laptop today.