Recursive releases Flint, a free laptop coding model it says AI largely built

Recursive, an AI startup founded last year that is trying to build AI that can improve itself, released its first model, Flint, on Wednesday. TechCrunch describes Flint as a free coding model that can run on a laptop and was built largely with the help of AI.

The AI in question is Recursive's automated research system, which it calls the Eureka Machine. The system comes up with ideas, runs experiments, and uses the results to decide what to try next. The company says it helped with nearly every step of Flint's development, and according to the launch video it contributed to nearly every part of the build, from pretraining to post-training and inference. That let a small team build the model from scratch in under four months, per the company. Recursive said Eureka's contributions include a sparser model shape, low-bit training that shrank Flint by 71% with almost no loss in quality, and custom kernels, which are specialized programs that speed up the underlying computations. Recursive did not say how much time the Eureka Machine saved its researchers.

Flint is a mixture-of-experts model: it uses only 2.5 billion of its 40 billion parameters at a time. That keeps it light enough for an ordinary laptop, and far smaller than Meta's Muse Spark Open, an open model built for Macs released last week, which uses 816 billion parameters and needs at least a terabyte of memory. Because Flint runs locally, developers can use it offline, keep their code private, and avoid the usage fees cloud AI services charge based on how much text the model processes. It is free to download on Hugging Face under the Apache 2.0 license, an open source license that allows free use and modification.

On speed, Recursive's own tests had Flint generating about 131 tokens per second on a MacBook Pro with Apple's M4 Pro chip. That is 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. On SWE-Bench Verified, a benchmark of how well AI fixes real-world software bugs, Flint outperformed other models under 50 billion parameters. On Recursive's own average across six tests, only one model, Alibaba's Qwen3.8 27B, scored higher.

The company acknowledged that Flint would need more guidance than the big paid coding assistants and suggested breaking large tasks into smaller steps. 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.

Flint is framed as an early public test of recursive self-improvement, the idea that AI can build better AI, which builds better AI still. It is one of the most debated ideas in the industry, and some researchers fear that AI improving itself could slip out of human control. For now, Recursive's version is more limited: an AI research assistant overseen by people that produced a model small enough for a laptop. Recursive says each new model should make the Eureka Machine a better researcher, helping it build the next one faster. Socher told Wired the longer-term goal is full recursive self-improvement, with humans overseeing the process and able to stop it if something goes wrong.

Background: Recursive emerged from stealth in May with $650 million in funding at a $4.65 billion valuation, from backers including GV, Nvidia, and AMD Ventures. It is led by CEO Richard Socher, Salesforce's former chief scientist. In June it signed a $410 million compute deal with Amazon Web Services. TechCrunch also notes that Recursive is 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.

Key facts

Flint is a mixture-of-experts coding model with 40 billion parameters, 2.5 billion active at a time, free to download on Hugging Face under Apache 2.0.
Recursive says its Eureka Machine research system helped with nearly every step, from pretraining to post-training and inference, letting a small team build Flint in under four months.
In Recursive's own tests Flint ran at about 131 tokens per second on a MacBook Pro with M4 Pro, versus 58 to 69 for similar models, and beat other models under 50 billion parameters on SWE-Bench Verified.
Wired's Will Knight, on a Mac with 48GB of memory, said Flint handled all his coding requests, though the company admits it needs more guidance than big paid assistants.
Recursive's self-improvement is limited for now: an AI research assistant overseen by people; the stated long-term goal is full recursive self-improvement with human oversight.

Why it matters

Flint is an early public test of recursive self-improvement, the idea that AI builds better AI, which is among the most debated ideas in the industry. Recursive says its research system took part in nearly every stage of building a model that now runs on a laptop. It is also small next to Meta's Muse Spark Open, which has 816 billion parameters and needs at least a terabyte of memory, against Flint's 2.5 billion active of 40 billion total. The company says each new model should make the Eureka Machine a better researcher and speed up the next one.

Who it affects

Developers who want a coding model on their own machine are the main audience: it works offline, keeps code private, and avoids per-use fees charged by cloud AI services. It also touches the wider debate over AI that improves itself, where some researchers fear it could slip out of human control, and rival makers of small open models such as Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning, which Recursive compared against.

How to use it

Download Flint free from Hugging Face; it is released under the Apache 2.0 license, which allows free use and modification. Recursive suggests breaking large tasks into smaller steps, since Flint needs more guidance than the big paid coding assistants. Will Knight of Wired ran it on a Mac with 48GB of memory to build small programs such as a to-do list app and a web research tool. The source does not state minimum hardware requirements.

How solid is it

The speed and benchmark figures come from Recursive's own tests, and the 71% size reduction and the Eureka Machine's role in building Flint are the company's account. The only outside check mentioned is Wired's Will Knight, whose hands-on impressions were positive. The open weights mean others can test the model. No SWE-Bench Verified score is given as a number, and Flint's own score on the six-test average is not given; only that Alibaba's Qwen3.8 27B scored higher. Recursive did not say how much time the Eureka Machine saved, and the source describes the AI's contribution only as 'largely' and 'nearly every step'.

Risks and caveats

Flint will need more guidance than big paid coding assistants, by the company's own admission. Some researchers fear that AI improving itself could slip out of human control; Recursive's current version is overseen by people, and Socher told Wired that the long-term goal includes humans able to stop the process if something goes wrong. The performance claims are the company's, and the amount of work the AI actually did is not quantified.

“less about headcount and more about agent count.”

— Recursive, describing its $410 million compute deal with Amazon Web Services, in a previous statement to TechCrunch