The 27b model inference speed is amazing at 35 tok/s with power limit to 260W. My cpu is 5950x, maybe that matters too. Really enjoying it and it’s my default now.