3 ms·

I really hope someone can come up with fast local inference hardware good enough to stop using cloud providers.