7 ms·

I feel like even TPS is becoming less of a good metric as we're seeing certain models handle similar problems while burning far fewer tokens.