Here is another test with Llama 4 Scout - 6 bit quantized version on M4 Max. Got ~30t/s which is faster than most of the hosted models.