This doesn't seem to take into account CPU MoE, which can make a huge amount of difference - Running a bigger MoE model is better than a small model that fits in your GPU if you have the CPU resources. I run Qwen3.6 (the 30b/e4b version) in MoE at around 40 tok/s on my 5070+Ryzen 9 5950X, and it's way better than that tool's recommended 9b.