Recursive launches Flint, a laptop-sized coding model its AI built

Recursive, an AI startup founded last year, released its first model on Wednesday: Flint, a free coding model that runs on an ordinary laptop and was, by the company's own account, built largely by AI. The startup's automated research system, called the Eureka Machine, helped with nearly every step of Flint's development. That let a small team build the model from scratch in under four months.

The release matters beyond the product itself. Flint is an early public test of recursive self-improvement — the idea that AI can build better AI, which in turn builds even better AI. It's one of the most debated concepts in the industry, and some researchers fear that AI improving itself could slip out of human control. Recursive's current version is more limited: an AI research assistant overseen by people, which produced a model small enough for a laptop.

Who is behind Flint?

Recursive emerged from stealth in May with $650 million in funding at a $4.65 billion valuation. Its backers include GV, Nvidia, and AMD Ventures. The company is led by CEO Richard Socher, Salesforce's former chief scientist.

The company is also serious about compute. In June, it signed a $410 million compute deal with Amazon Web Services. Recursive previously told TechCrunch that the deal was "less about headcount and more about agent count." That one line explains the whole business model: Recursive wants machines doing the research work, not a growing payroll of humans.

Why run a coding model on a laptop?

Because Flint runs on a user's own computer rather than a company's cloud servers, developers can use it offline, keep their code private, and avoid the usage fees that AI cloud services charge based on how much text the model processes. Those are concrete advantages for anyone working with sensitive code or in places without reliable connectivity.

The model is free to download on Hugging Face, the popular site for sharing AI models, under the Apache 2.0 license. That's an open-source license that allows free use and modification, so other developers can inspect, change, and redistribute Flint without asking permission.

How small is small?

Flint is a mixture-of-experts model, meaning it uses only a small slice of itself for any given task: just 2.5 billion of its 40 billion settings, known as parameters, at a time. That design choice is what makes it light enough for an ordinary laptop.

The comparison with the competition is stark. Meta's Muse Spark Open, an open model built for Macs released last week, uses 816 billion parameters and needs at least a terabyte of memory. Flint fits in a fraction of that footprint, which is the entire point: it trades away the scale of datacenter models for portability and cost.

How fast is it?

In Recursive's tests, Flint generated about 131 tokens — chunks of text — per second on a MacBook Pro with Apple's M4 Pro chip. That's roughly twice the 58 to 69 tokens per second that similar models managed on the same machine, including Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning. On Nvidia's RTX Spark, Flint topped 150 tokens per second.

Speed matters for coding assistants because developers interact with them conversationally, iterating on code in real time. A model that takes seconds per line quickly becomes a model nobody uses.

How well does it actually code?

On SWE-Bench Verified, a benchmark that measures how well AI can fix real-world software bugs, Flint outperformed other models under 50 billion parameters. On Recursive's own average across six tests, only one model scored higher: Alibaba's Qwen3.8 27B.

The company acknowledged that Flint would need more guidance than the big paid coding assistants, and suggested breaking large tasks into smaller steps. Still, early reviews were positive. Wired's Will Knight used Flint on a Mac with 48GB of memory to build several small programs, including a to-do list app and a tool that scours the web for research. He found that smaller coding models often ran into errors or were slow compared to Claude or Codex, but Flint nailed his coding requests every time.

That review, from an experienced observer rather than the company, is arguably the strongest signal in the whole launch.

What did the Eureka Machine actually do?

The Eureka Machine comes up with ideas, runs experiments, and uses the results to decide what to try next. According to the company's launch video, it contributed to nearly every part of building Flint: pre-training, post-training, and inference.

Recursive said Eureka's contributions include:

• A sparser model shape, which reduces how much of the model activates at once
• Low-bit training that shrank Flint by 71% with almost no loss in quality
• Custom kernels — specialized programs that speed up the underlying computations

The 71% shrinkage figure is the headline number here. Compressing a model that much while keeping quality nearly intact is exactly the kind of optimization work that normally eats months of researcher time.

Recursive didn't say how much time the Eureka Machine saved its researchers. But it's not alone in leaning on AI to do 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 industry-wide direction is clear: labs are increasingly using their own models as research staff.

Is this really recursive self-improvement?

Not yet, and Recursive is fairly candid about that. The current setup is an AI research assistant with people in the loop, not an autonomous scientist. The longer-term goal, Socher told Wired, is for the Eureka Machine to achieve full recursive self-improvement, with humans overseeing the process and able to stop it if something goes wrong.

The feedback loop Recursive describes is straightforward: each new model should make the Eureka Machine a better researcher, helping it build the next one faster. If that bet pays off, each new model should be able to take on longer stretches of work without human help.

That's the experiment Flint quietly starts. A free, laptop-sized coding model is the visible product; the real test is whether the machine that built it can build the next one better.