Recursive has released its first model, and it is not asking for your API key. Flint is an open-weight coding model designed to run on your own machine, and the company says it pushes past 125 tokens per second on a MacBook Pro.
What Recursive actually shipped
Flint launched on October 21, 2026. It is a 40-billion-parameter model built specifically for coding work.
On a MacBook Pro with an M4 Pro chip, Flint generates over 125 tokens per second. On RTX Spark hardware, Flint could potentially reach 150 tokens per second or more. Recursive’s performance claims say the model can beat comparable models by up to two times.
For agentic coding, where the model takes multi-step actions rather than simply autocompleting a line, Flint is described as workable on consumer machines with 48GB of RAM.
Flint is also positioned among the leading Mixture of Experts models under 50 billion parameters, particularly for agentic coding. A Mixture of Experts design works a bit like a hospital triage desk: instead of every specialist seeing every patient, the system routes each request to the parts of the model best suited to handle it. The payoff is typically faster responses without running the whole network on every step.
A model built mostly by machines
Recursive describes itself as an AI startup focused on recursive self-improvement, or RSI. The idea is that AI systems help design and improve the next generation of AI systems.
Flint is the first public product of that approach. A small team built it from scratch in less than four months, leaning heavily on a proprietary automated research system the company calls the Eureka Machine.
Recursive’s broader goal is to automate multiple stages of model development, including tuning architecture and refining training methods.
The money behind the machine
Recursive has raised $650 million at a valuation of $4.65 billion. Its backers include Google Ventures, Greycroft, Nvidia, and AMD Ventures.
Why local and open matter here
Flint’s weights are available on Hugging Face. Open weights mean anyone can download the trained model and run it themselves, free of charge, without routing code through a third-party server.
Cloud coding assistants send your code to an outside provider, and many companies are uneasy about proprietary source code leaving their network. A local model keeps everything on the machine. Flint is pitched at local, unmetered use, which suits the kind of developer who leaves an agent running through a long refactoring job and would rather not check the invoice afterward.
What this means for developers and the AI market
There are reasons for caution. Performance claims like “up to two times” faster come from the company, and independent testing across varied hardware and workloads will show how well Flint holds up outside launch-day conditions. The 150 tokens per second figure on RTX Spark hardware is also framed as a potential ceiling, not a guarantee.
Hardware is the other constraint. A 48GB machine sits at the premium end of the consumer market, so “local” still means “local, if you have a well-equipped computer.”
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.