I usually get 7 token/s using llama or lm studio, but this inference recipe runs at a smooth 80 tokens per second.

Genuinely very usable, and fully local!