$ onpar measure --endpoint http://127.0.0.1:8001/v1 --served-model Qwen/Qwen3-30B-A3B-GGUF:Q8_0 --model qwen3-30b-a3b --quant q8 --samples 3


  Qwen/Qwen3-30B-A3B-GGUF:Q8_0 via http://127.0.0.1:8001/v1

    sample 1     91.78 tok/s decode   ttft    122 ms   127 tokens
    sample 2     92.07 tok/s decode   ttft     15 ms   127 tokens
    sample 3     91.69 tok/s decode   ttft     15 ms   127 tokens

    median     91.78 tok/s   spread 0.4%
    roofline  118.59 tok/s   estimate
    measured is 77% of the estimate

    roofline basis: --quant q8; at the 152-token context the samples ran at (prompt + output) — pass --context to compare against a ceiling instead; measured implies 56% of theoretical bandwidth, against the 72% the planner assumes

    before: load 8.96 over 16 cores (0.56/core) · busiest: mediaanalysisd 154%, perl 100%, Python 83%
    after:  load 10.09 over 16 cores (0.63/core) · busiest: mediaanalysisd 104%, perl 100%, WindowServer 45%

  Usable. Still one machine, one moment, one prompt shape —
  report it as measured here, not as this model's speed.



# recorded 2026-11-03 on M4 Max · 128 GB · onpar 1.5.2
