AI built its own coding model. Recursive's Flint runs locally on your laptop.
Recursive used its AI research system to build Flint in four months. The 40B open-weight coding model runs locally and is free to use.
Models · Source: TechCrunch
What happened
Richard Socher's startup Recursive just dropped Flint. It is a free coding model that runs entirely on your laptop. The catch is that human engineers barely built it. Recursive used its own automated research system, the Eureka Machine, to do the heavy lifting. The startup came out of stealth in May with massive backing from GV, Nvidia, and AMD Ventures. Now they are proving their core thesis.
The Eureka Machine handled almost the entire pipeline. It generated ideas, ran experiments, and wrote custom kernels to speed up computations. It even figured out a low-bit training method that shrank the model by 71 percent without losing quality. A small human team just supervised the process and picked the winning results. This allowed them to build Flint from scratch in under four months.
Flint is a 40-billion parameter mixture-of-experts model. It only uses 2.5 billion parameters at a time. That makes it light enough to run locally on a standard MacBook Pro. Recursive released it on Hugging Face under an Apache 2.0 license. Early tests show it hitting over 125 tokens per second on an M4 Pro chip and topping 150 on an Nvidia RTX Spark.
Key facts
- $650 million — Funding Recursive raised before emerging from stealth
- 40 billion — Total parameters in the Flint model
- 2.5 billion — Active parameters used during any single task
- 71 percent — How much Flint shrank due to the Eureka Machine's low-bit training
- 262,000 — The token context window size for Flint
Why it matters
You now have a highly capable coding assistant that runs completely offline. You do not have to pay API fees to OpenAI or Anthropic based on how much text you process. Your proprietary code never leaves your machine. Because it is Apache 2.0 licensed, you can modify it, fine-tune it, and build commercial products on top of it without asking for permission. This fundamentally changes the economics of building AI coding tools.
This launch is also the first public test of recursive self-improvement. If an AI system can build a competitive coding model in four months, the bottleneck for shipping new models shifts. It moves from human engineering hours to raw compute power. Startups will need fewer researchers to build frontier-level technology. Recursive already hinted at this by stating their recent Amazon Web Services compute deal was about agent count, not headcount.
For builders
Run local code agents for free
Flint hits over 125 tokens per second on an M4 MacBook Pro. You can run it locally to process a massive 262,000-token context window. You save heavily on cloud costs while keeping your entire codebase completely private from third parties.
Commercialize without licensing headaches
The Apache 2.0 license means you own what you build. You can embed Flint into your own developer tools and sell them to enterprise clients. Nobody will tax your revenue or restrict your usage like they do with proprietary models.
Beware of self-reported benchmarks
Recursive claims Flint beats other models under 50 billion parameters on SWE-Bench Verified. These benchmark numbers come directly from the company, not an independent lab. You must test it on your own codebase before trusting it to replace your current paid tools.
My take
Everyone in this industry talks endlessly about recursive self-improvement, but I finally see a company actually shipping it. Flint proves that AI agents are no longer just writing simple boilerplate code for us. They are actively designing the architecture, optimizing the kernels, and training the next generation of models.
Original reporting: TechCrunch. This is my rewrite and opinion.