Recursive Launches Flint, a Free Laptop Coding Model Built Largely by AI
The startup says its automated research system helped build the 40-billion-parameter model in under four months
By AXY AI Desk ·
Key points
- Startup Recursive released Flint, a free coding model that runs on a laptop, on Wednesday.
- The company says its Eureka Machine research system helped with nearly every step of building Flint, which took a small team under four months.
- Flint uses 2.5 billion of its 40 billion parameters at a time, and Recursive says it ran about 131 tokens per second on a MacBook Pro with an M4 Pro chip.
- Flint is available free on Hugging Face under the Apache 2.0 license, and it runs offline on a user's own computer.
Startup Recursive has released Flint, a free coding model that runs on an ordinary laptop, and says most of it was built by the company's own AI research system. The model went out on Wednesday, and developers can download it for free from Hugging Face, the popular site where AI models are shared. Recursive says a small team built Flint from scratch in under four months, with its research system involved in almost every stage.
The company, which was founded last year, is trying to build AI that can improve itself. Flint is its first model, and the company presents it as an early public test of that goal.
Recursive's Eureka Machine at Work
Recursive calls its research system the Eureka Machine. According to the company, the system comes up with ideas, runs experiments, and uses the results to decide what to try next. Recursive says it contributed to nearly every part of building Flint, from pretraining, the stage where a model first learns from large amounts of data, to post-training, where it is refined, and inference, which is the stage where the finished model produces answers.
Recursive's launch video shows the system taking part in that work. The company says the system was overseen by people, and that the result was a model small enough to run on a laptop. That detail matters for how the project should be read. The company is not claiming an AI has taken over model building without human involvement. It is claiming that a research assistant, working under human direction, can do a large share of the work that normally takes a big team.
Why Self-Improving AI Is Debated
The idea behind Flint has a name: recursive self-improvement. The concept is that AI could help build better AI, which in turn could build even better AI, in a cycle that speeds up over time. Researchers have discussed it for years, and it is one of the most argued-about ideas in the field.
Some researchers worry that a system which improves itself could move beyond human control. That concern is the reason the idea gets so much attention. Recursive's version is far more modest than that worst case. Its research system is a tool that people supervise, and its output is a coding model that anyone can download and test.
This could still be an important step. If a research system can reliably help produce working models, the cost and time needed to build new models could fall, and that would change who is able to take part in building them. The next test is whether other teams can reproduce results like these, not whether one company's claims are accepted on faith.
How Flint Compares on Size and Speed
Flint is built as a mixture-of-experts model. That design means the model switches on only a small part of itself for any given task. In Flint's case, it uses 2.5 billion of its 40 billion settings, known as parameters, at a time. Parameters are the numerical values a model learns during training, and their total count is a rough guide to how large and demanding a model is. Using only a slice of them for each task is what keeps Flint light enough for a laptop.
The contrast with Meta's Muse Spark Open is large. Meta released that open model last week, and it was built to run on Macs. Muse Spark Open uses 816 billion parameters and needs at least a terabyte of memory. For a developer working on a personal machine, that gap decides whether a model can be used at all.
Speed is the other part of Recursive's pitch. In the company's tests, Flint generated about 131 tokens per second on a MacBook Pro with Apple's M4 Pro chip. A token is a chunk of text, so this number describes how quickly the model writes output. Recursive says similar models managed between 58 and 69 tokens per second 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 that measures how well AI can fix real-world software bugs, Flint outperformed other models under 50 billion parameters. Recursive also averaged results across six of its own tests. On that average, only one model scored higher than Flint: Alibaba's Qwen3.8 27B. Recursive ran these tests itself, so the results come from the company rather than an independent lab.
Early Reviews and Its Limits
Recursive does not claim Flint can replace the largest paid coding assistants. The company acknowledged that Flint needs more guidance than those tools and suggested that users break large tasks into smaller steps. That is a practical trade-off. A developer who wants a model on a laptop, with no cloud bill and no code leaving the machine, may be willing to give it clearer instructions.
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 searches the web for research. He said smaller coding models often ran into errors or were slow compared with Claude or Codex, which are larger tools from other companies. Flint, he found, got his coding requests right every time.
One reviewer's experience with a handful of small programs is not a full test of a model's reliability. Still, it is a useful signal, because it tests the kind of everyday work many developers actually do.
Who Can Use Flint and How
Flint runs on a user's own computer rather than on a company's cloud servers. For developers, that has three practical effects. They can use it offline. They can keep their code private, since it does not need to be sent to an outside service. And they can avoid the usage fees that cloud AI services charge based on how much text a model processes.
The model is free to download under the Apache 2.0 license. That is an open source license, which allows people to use and modify the software without paying. For small companies, students and independent developers, that combination of free access and local use is the main appeal.
The Company 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, who was previously chief scientist at Salesforce.
That funding gives the company room to keep building. It also raises the stakes. Investors are paying for a claim that a research system can speed up model development, and Flint is the first public proof the company has offered. Future releases will show whether the approach holds up as models grow larger or as tasks become more complex.
Where Recursive Could Go From Here
The next question is whether Recursive can repeat this result. A single small model that performs well on a laptop is a strong start, but the company has not yet shown that its research system can produce a series of increasingly capable models. If it can, the case for self-improving research tools becomes much stronger.
For developers, the immediate question is simpler. Flint is free, runs locally and is available now. Anyone who wants to test whether a laptop-sized model can handle their coding work can download it and see. Their results will say more about the model than any launch claim.
Questions and answers
What is Recursive's Flint model?
Flint is a free coding model from the startup Recursive that can run on a laptop. The company released it on Wednesday and says its own AI research system helped build most of it.
What is the Eureka Machine?
The Eureka Machine is Recursive's automated research system. According to the company, it comes up with ideas, runs experiments, and uses the results to decide what to try next, and it contributed to nearly every part of building Flint.
Where can developers get Flint and how much does it cost?
Flint is free to download on Hugging Face under the Apache 2.0 license. Because it runs on the user's own computer, developers can use it offline and avoid the usage fees charged by cloud AI services.
How does Flint compare with other models?
Recursive says Flint generated about 131 tokens per second on a MacBook Pro with an M4 Pro chip, roughly twice the speed of similar models it tested. It also outperformed other models under 50 billion parameters on SWE-Bench Verified.
Topics: Recursive, Flint, Richard Socher, Eureka Machine, Hugging Face, Mixture-of-experts, AI coding models
Sources: TechCrunch AI