A laptop can run it, the tests are theirs

This TechCrunch page is stamped 1:27 PM PDT · October 21, 2026. The judgment first: Recursive released its first model, Flint, on Wednesday. Free, able to run on a laptop, and built largely with the help of AI are the company's statements. About 131 tokens per second is the company's test on one MacBook Pro, not an independent check printed here. On SWE-Bench Verified the page says Flint outperformed other models under 50 billion parameters. The score is not printed.

Flint is a mixture-of-experts model. Any one task uses only a small slice of it: 2.5 billion of its 40 billion parameters. The page says that makes it light enough for an ordinary laptop. The comparison named here is Meta's Muse Spark Open, released last week, an open model with 816 billion parameters that needs at least a terabyte of memory. The size gap is the comparison on this page, not a win on one shared test. The license is Apache 2.0, and the download on Hugging Face is free.

The speed is also the company's own test. On a MacBook Pro with Apple's M4 Pro chip, Flint generated about 131 tokens per second, roughly twice what similar models managed on the same machine. Those models include Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning, at about 58 to 69 tokens per second. On Nvidia's RTX Spark, Flint topped 150 tokens per second. The machines and the token measure are what the page states. What would change on another machine is not on this page.

The bug-fixing rank also has no score. The page says that on SWE-Bench Verified, a benchmark of fixing real software bugs, Flint outperformed other models under 50 billion parameters. On the company's own average across six tests, only Alibaba's Qwen3.8 27B scored higher than Flint. Which six tests, and what the average was, are not printed. The company also acknowledged that Flint would need more guidance than the big paid coding assistants, and suggested breaking large tasks into smaller steps. More guidance is the company's statement, not a failure rate.

Built largely with AI refers to the automated research system the company calls the Eureka Machine. The page says it contributed to nearly every part of building Flint, from pre-training to post-training and inference, and that a small team built the model from scratch in under four months. The company says its contributions include a sparser model shape and low-bit training that shrank Flint by 71 percent with almost no loss in quality. The 71 percent is a result the company claims. This page does not set out the training steps, and this piece does not add them. Custom kernels are described only as specialized programs that speed up the computations. The method is not written here.

Wired's Will Knight used Flint on a Mac with 48GB of memory to build several small programs, including a to-do list and a tool that searches the web for research. His report is that smaller coding models often ran into errors or were slow, while Flint completed his coding requests every time. That is one journalist's trial, not a benchmark table. The $650 million raised at a $4.65 billion valuation when the company left stealth in May, and the $410 million compute deal with Amazon Web Services in June, are earlier background. Full recursive self-improvement, which CEO Richard Socher described to Wired, is a longer-term goal. The page says humans could stop the process if something went wrong. That is a goal, not something that happened on Wednesday.