NVIDIA's Open Coding Model Runs on Your Desk. The Cloud Just Lost a Job.
Recursive Flint is a 40B open model that codes locally at 220 tokens per second. That's a line in the sand.
Alongside the hardware, NVIDIA introduced Recursive Flint, a 40-billion-parameter mixture-of-experts coding model with open weights under Apache 2.0, reported above 220 tokens per second on RTX Spark hardware.
A model that good, that small, that local means a lot of 'send it to the API' coding tasks can stay on the machine. Latency drops, and your code never leaves the box.
It won't replace the frontier model for the hard reasoning jobs. But the routine 80 percent -- boilerplate, tests, refactors -- now has a free, private option.
If you build dev tooling, plan for local models doing the boring parts. The cloud stops being the only place code gets written.