3 ms·

I wouldn't go lower than Q3_K_S, as its basically the same filesize, and llama 33B has a big perplexity dropoff.