Recursive put Flint out on October 21 with a claim that cuts through the usual GPU theater. A coding model, open weights, and a stated speed above 125 tokens a second on a MacBook Pro.
The size, read carefully
Forty billion parameters sounds like a data-center model. Flint is a mixture of experts, and Recursive says only about 2.5 billion of them are active on a given step. That is the whole trick behind the laptop sentence. Context goes to 262,000 tokens, which is enough for a serious repository if you are disciplined about what you paste. The license is Apache 2.0.
The origin story is part of the release. Recursive says a small team built Flint in under four months, using an automated research system. Take that as a description of their process, not as proof that anyone with a script can repeat it.
Try it on your repo, quietly
A local coder changes the privacy math. The file does not have to leave the machine. It also changes the quality math, because a small active model will miss things a hosted frontier model catches. Run it on a bug you already understand before you run it on a bug you do not.
The 125-token figure is theirs, on an M4 Pro. If your laptop is older, assume less and measure. Marketing speed and your fan curve are not the same instrument.