TechCrunch reported that AI startup Recursive released its first model, Flint, on Wednesday. Recursive said Flint has 40 billion parameters (the settings a model learns) but uses only 2.5 billion for each token, a small chunk of text. The company said its research system, the Eureka Machine, helped build nearly every part of the model. Recursive said a small team finished it in under four months. The weights are free on Hugging Face under the Apache 2.0 license. Recursive said Flint generates over 125 tokens per second on an M4 Pro MacBook Pro. TechCrunch put it at about 131 in Recursive's tests, roughly twice the speed of similar models.
The speed and benchmark figures come from Recursive's own tests. TechCrunch said that on Recursive's average across six tests, only Alibaba's Qwen3.8 27B scored higher. Recursive said Flint may need more guidance than larger closed-source coding models. It advised splitting big jobs into smaller steps. TechCrunch said Wired's Will Knight built several small programs with Flint on a 48GB Mac, and it handled his requests every time. The sources show no other independent testing.
Flint runs on the user's own machine. Recursive says that means teams can work offline, keep code private, and avoid per-use fees. Teams weighing local coding tools can now test a free, modifiable option. Recursive's own advice is that it needs more guidance than larger paid assistants.