>>107119404
>>107119586
>>107120393
Ok, ran the test. If Dave2D's test for the Studio is accurate (https://www.youtube.com/watch?v=J4qwuCXyAcU) then the Pro 6000 is simply slower than the M3 Ultra for running large MoEs like Deepseek R1 (16 tk/s vs 14 tk/s at Q4).
For more medium size models like Qwen or GLM I doubt there would be much difference since the number of active parameters is similar. And the build would be ~50% more expensive.