A startup built a coding model that scores 74.2% on SWE-bench Verified, runs locally on your MacBook, and used AI to build most of itself. Recursive launched Flint on October 21, 2026. They shipped it in under four months with a small human and agent team. That timeline should make every major AI lab uncomfortable.
Why This Launch Matters Right Now
Most coding models live in the cloud. You send code up, pay per token, and hope your proprietary logic does not end up in someone’s training data. Flint breaks that model. Its weights are free to download under the Apache 2.0 license, according to TechCrunch, and it runs on consumer hardware including the MacBook Pro M4 Pro and NVIDIA RTX GPUs.
The timing is not random. NVIDIA announced on the same day that its RTX Spark, DGX Spark, GeForce RTX, and RTX PRO GPUs all support Flint at launch, according to NVIDIA’s blog. DGX Spark 64GB units become available October 23. That is a coordinated hardware and software push, and it signals that local AI inference is being treated as a serious market, not a hobby niche.
Recursive emerged from stealth in May 2026 with $650 million in funding at a $4.65 billion valuation, according to TechCrunch. Backers include GV, NVIDIA, and AMD Ventures. CEO Richard Socher previously served as Salesforce’s chief scientist. In June, Recursive signed a $410 million agreement with Amazon Web Services, according to TechCrunch. This is not a scrappy indie project. The resources behind Flint are substantial.
The Part Nobody Is Talking About
Here is what actually matters. Recursive says its internal system called the Eureka Machine helped with nearly every stage of Flint’s development, according to TechCrunch. The system proposes ideas, runs experiments, and uses results to guide the next round of research. A small human and agent team built a competitive coding model from scratch in under four months using this approach.
I want to be clear here. “AI built it” is Recursive’s claim, not an independently verified fact. Recursive has not published measured researcher time savings or disclosed how much human oversight went into each decision. What we know is that this is human supervised research automation. It is not a robot that woke up and wrote its own weights in a dark server room. But even the supervised version of what they describe represents a real shift in how AI labs can operate.
Think about what that four month timeline means for the economics of this industry. Traditional lab math assumed you needed hundreds of researchers, billions in compute, and years of iteration to build frontier models. Recursive’s strategy, in their own words, is “less about headcount and more about agent count,” according to TechCrunch. If that framing holds, the cost structure of building competitive AI models just changed.
The technical architecture supports this efficiency story. Flint uses a mixture of experts design with approximately 40 billion total parameters but only 2.5 billion active per token, according to TechCrunch and LM Studio documentation. That means it does not process the full model for every word it generates. LM Studio documents a context window of 262,144 tokens. On a MacBook Pro with an M4 Pro chip, Recursive reports approximately 131 tokens per second compared to 58 to 69 tokens per second for comparison models tested on the same hardware, according to TechCrunch. That speed gap matters in real workflows.
NVIDIA’s involvement goes beyond distribution. NVIDIA says its Eureka generated specialized kernels increased Flint’s generation speed on RTX Spark by 55% without modifying the model’s weights, according to NVIDIA’s blog. That is optimization at the deployment layer, not the training layer. A third party confirmed measurable gains from this approach on their own hardware.
The benchmark numbers are company reported: 74.2% on SWE-bench Verified, 89.7% on SWE-bench Pro v2, and 54.9% on Terminal Bench 2.1, per Recursive’s own disclosures as reported by TechCrunch. They have not been independently verified. Treat them as performance claims, not settled science. But even as claims, they are competitive with models that cost significantly more to access.
For developers who want to document or share their workflow experiments, InVideo AI makes it easy to turn screen recordings and notes into short explainer videos without a production team. Not every insight needs to stay buried in a Notion doc.
What This Means for You
If you write code for a living, Flint is worth testing today. Not because it is proven to be the best model, but because the price is zero and the data stays on your machine. For any developer working with proprietary codebases, client work, or anything you would not want going to a third party server, local inference is not optional. It is basic operational security.
Here is what I would do. Download Flint through LM Studio this week. Run it on a real task from your current project. Not a toy benchmark. An actual problem you are stuck on. Compare the output quality to whatever cloud model you are paying for now. If the quality is close enough and the speed holds, you just eliminated a recurring line item from your budget.
For small teams and solo developers, the cost savings are real in one specific way. No token charges during local inference. But do not assume it replaces cloud models for every use case. Some tasks still favor scale and fine tuning that runs on larger infrastructure. The honest answer is that Flint is worth adding to your toolkit even if you do not use it exclusively.
For entrepreneurs and operators watching this space, the Eureka Machine story is the bigger signal. If Recursive can repeat this four month build cycle and keep improving the quality of what the system produces, the cost to build competitive AI products will keep falling. That creates real opportunity for smaller teams to ship things that would have required a well funded lab two years ago. AppSumo regularly surfaces lifetime deals on developer tools built by exactly these kinds of small teams moving fast. Worth checking what shows up there over the next few months as this wave accelerates.
The Bottom Line
Recursive shipped a competitive coding model in four months using AI to help build it, licensed it for free, and got NVIDIA hardware backing on day one. Every developer and small team should download it and stress test it against real work. The free price is not the point. The point is that building competitive AI tools just got cheaper and faster. The question is not whether this trend continues. It is whether you are positioned to use it before the people competing with you are.
Frequently Asked Questions
What is Flint by Recursive?
Flint is a coding model launched by Recursive on October 21, 2026. Its weights are free to download under the Apache 2.0 license and it runs locally on supported laptops and NVIDIA hardware. It uses a mixture of experts design with roughly 40 billion total parameters and 2.5 billion active per token, according to TechCrunch and LM Studio.
Did AI really build Flint?
Recursive says its Eureka Machine helped with nearly every stage of development, according to TechCrunch. A small human and agent team completed the build in under four months using human supervised automation. Recursive has not published independent verification of exactly how much work was AI generated versus human directed, so treat the “built by AI” framing as an attributed company claim.
How fast does Flint run on a MacBook?
According to Recursive’s own testing as reported by TechCrunch, Flint runs at approximately 131 tokens per second on a MacBook Pro with an M4 Pro chip. That compares to 58 to 69 tokens per second for comparison models tested on the same machine. These are vendor reported figures, not third party benchmarks.
Is Flint free to use commercially?
Yes. The model weights are released under the Apache 2.0 license, meaning you can download and use them commercially without paying Recursive. You pay for the hardware and electricity to run inference locally, but there are no per token charges during local use.
What does the Recursive funding level tell us about this space?
Recursive raised $650 million at a $4.65 billion valuation in May 2026, according to TechCrunch, with backers including GV, NVIDIA, and AMD Ventures. A $410 million AWS compute agreement followed in June. That scale of backing from hardware companies specifically suggests the investors believe AI assisted model development can produce commercially competitive results at lower cost than traditional lab methods.
