3 ms·

I don't see any special sauce mentioned to make it run these models fast. If you get a token per second this thing is useless