Skip to content
llm-speed
← raw run
wrapped·Jun 27, 2026
/r/r_roktphpc--8
your run

qwen3-coder-bench-32k

on M4 Max
113
tok/sdecode
M4 Max (40-core GPU) + 128GB unified
rank in tier
1/5M4 Max runstop 20%
best workload
chat-short
where the rig flew
slowest workload
chat-long97.4 tok/s
where the rig struggled
backend
ollamaollama is llama.cpp under the hood; pure llama.cpp tends to nudge a few % faster.
faster than
  • granite-8b-code-instruct-4bit on M3 Ultra112 tok/s
  • Qwen3-Coder-30B-A3B-Instruct-4bit on M3 Ultra112 tok/s
  • Yi-Coder-9B-Chat-4bit on M3 Ultra104 tok/s
share
113 tok/s on M4 Max running qwen3-coder-bench-32k — top 20% of M4 Max runs. Signed benchmark via llm-speed.
https://llm-speed.com/wrapped/r_roktphpc--8
X / TwitterBlueskyMastodonRedditraw run

This is a shareable view of one signed llm-speed suite-suite-v1 run. Numbers are not edited. The raw record has every workload, the public key, and the fingerprint hash.