I think it depends on the kind of answer and how long the round trip is. Even a fast model that waffles for several minutes before giving an actual answer (coug

I think it depends on the kind of answer and how long the round trip is. Even a fast model that waffles for several minutes before giving an actual answer (coughqwen3.5cough) can feel very slow. Few seconds per word output may be fine if the final answer is correct and short. But generally, I'd like to see above 20, >50 is mostly great, and more is better. For conversational response, that is, not batch or interactive loop.