Recursive releases Flint: 40B MoE open coding model that runs locally on a MacBook
Recursive's first release, Flint, is a 40B mixture-of-experts coding model with Apache 2.0 weights that runs entirely on a MacBook at 125+ tokens/s in short contexts, with no per-token bill.
An AI helped build a new coding AI in under four months, and the result runs on a MacBook at 125+ tokens a second at short contexts. It runs entirely on your own machine, so there's no bill per token. That model is Flint, the first release from Recursive, a 40B mixture-of-experts model with open weights under Apache 2.0. The research system behind it, the Eureka Machine, runs the scientific method in a loop, chasing lots of research paths in parallel and getting better at research from whatever comes back. Look at where its work showed up: architecture, low-bit training, specialized inference kernels and speculative decoding. If you sat down to make a 40B model fly on a laptop, that's the to-do list you'd write. The mixture-of-experts design only wakes up a slice of the model for each token, and the research system went after the levers that turn that into real speed. So when you watch it run on a MacBook, you're watching the research system's work. I'd love to do all my coding locally someday, and this is the closest that's felt. When the model lives on your machine you can leave an agent running all day without a meter ticking in the background, and your prompts and code stay on hardware you own. There's been a ton of hype around AI that improves itself, and I think we'll look back on this as one of the first real milestones: a small team and their research system shipped a model anyone can download and run on their own laptop.
This post quotes another post: x.com/1442906958/status/2112944385415426249