2 ms·

M5 is supposed to support FP4 natively which would explain the speed up on Q4 quantized models (down from BF16).