Here is llama-bench on the same M4:

  | model                    |       size |     params | backend    | threads |            test |                  t/s |
  | ------------------------ | ---------: | ---------: | ---------- | ------: | --------------: | -------------------: |
  | qwen35 27B Q4_K_M        |  15.65 GiB |    26.90 B | BLAS,MTL   |       4 |           pp512 |         61.31 ± 0.79 |
  | qwen35 27B Q4_K_M        |  15.65 GiB |    26.90 B | BLAS,MTL   |       4 |           tg128 |          5.52 ± 0.08 |
  | qwen35moe 35B.A3B Q3_K_M |  15.45 GiB |    34.66 B | BLAS,MTL   |       4 |           pp512 |        385.54 ± 2.70 |
  | qwen35moe 35B.A3B Q3_K_M |  15.45 GiB |    34.66 B | BLAS,MTL   |       4 |           tg128 |         26.75 ± 0.02 |

So ~60 for prefill and ~5 for output on 27B and about 5x on 35B-A3B.