AI startup Recursive released Flint on October 21, 2026, a free coding model designed to run on a laptop rather than in the cloud. The company says the model was built largely by its own automated research system in under four months, and it is available free on Hugging Face under the Apache 2.0 open source license.
What Flint is
Flint is a mixture-of-experts model with 40 billion total parameters, of which about 2.5 billion are active at any one time, according to Recursive’s own announcement. Because it runs locally on consumer hardware, developers can use it offline, keep their code private, and avoid ongoing cloud usage fees that come with hosted coding assistants.
The model is released under the Apache 2.0 license, which allows free commercial and personal use, modification, and redistribution.
How Recursive built it
Recursive is an AI startup founded in 2025 by Richard Socher, Salesforce’s former chief scientist, who serves as CEO. The company emerged from stealth in May 2026 with 650 million dollars in funding at a 4.65 billion dollar valuation, backed by GV, Nvidia, and AMD Ventures, per Recursive’s announcement.
According to the company, Flint’s development leaned heavily on Recursive’s internal automated research system, called the Eureka Machine, which generates research ideas, runs experiments, and uses the results to decide what to try next. Recursive says the system contributed a sparser model architecture and a low-bit training method that shrank Flint’s size by 71 percent with little loss in output quality, along with custom computing kernels.
Recursive describes its broader mission as building AI systems that can improve themselves, with Flint positioned as a public demonstration of that approach rather than a standalone product launch. The small team size and under-four-month timeline the company cites are meant to illustrate how much of the engineering work the Eureka Machine handled on its own.
How it performs, by Recursive’s own numbers
Recursive’s published benchmarks, which are the company’s own and have not been independently verified, show Flint running at about 131 tokens per second on a MacBook Pro with an Apple M4 Pro chip, compared with 58 to 69 tokens per second for similar-sized models it tested, including Gemma 4 and Nemotron 3.5 Lightning. On Nvidia’s RTX Spark hardware, the company reports speeds above 150 tokens per second.
On the SWE-Bench Verified coding benchmark, Recursive says Flint outperformed other models under 50 billion parameters. On the company’s own six-test average, it says only Alibaba’s Qwen3.8 27B model scored higher.
| Metric | Flint | Comparable models (Recursive’s tests) |
|---|---|---|
| Total parameters | 40 billion (2.5B active) | Varies |
| Speed on M4 Pro MacBook Pro | ~131 tokens/sec | 58 to 69 tokens/sec (Gemma 4, Nemotron 3.5 Lightning) |
| Speed on RTX Spark | 150+ tokens/sec | Not disclosed |
| License | Apache 2.0 (free) | Varies by model |
Why running locally matters for developers
Most widely used coding assistants run on company servers, which means a developer’s code, prompts, and project context pass through a third party’s infrastructure. For teams working on proprietary or regulated codebases, that can rule out using those tools altogether.
A model that runs entirely on a laptop removes that concern, since nothing leaves the machine. It also removes the recurring per-token or subscription costs that come with hosted coding assistants, since the only cost is the hardware the developer already owns. Recursive is positioning Flint as free infrastructure for individual developers and smaller teams who want those benefits without paying for compute.
Flint arrives in a market that already has several open-weight coding models, including Google’s Gemma line, Nvidia’s Nemotron series, and Alibaba’s Qwen models, all of which Recursive’s own benchmark comparisons name directly. What sets Flint apart, according to Recursive, is less about raw capability and more about how much of its own development pipeline, from model architecture to training compression, was handled by its automated Eureka Machine rather than human researchers.
What Recursive says Flint cannot do as well
Recursive itself says Flint needs more explicit guidance than larger, paid coding assistants, and it recommends breaking big coding tasks into smaller steps rather than handing the model an open-ended project. The company has not claimed Flint matches the largest commercial coding models on every task, and none of its benchmark figures have been independently verified by a third party as of publication.
Frequently asked questions
Is Flint free to use?
Yes. It is released on Hugging Face under the Apache 2.0 open source license, which allows free personal and commercial use.
Does Flint need an internet connection?
No. It is designed to run locally on a laptop, so it can work offline once downloaded.
Who built Flint?
Recursive, an AI startup led by CEO Richard Socher, Salesforce’s former chief scientist, which emerged from stealth in May 2026 backed by GV, Nvidia, and AMD Ventures.
Leave a Reply