wrapped·Nov 06, 2026
/r/r_-jalsgvd-bryour run
Qwen3.8-27B
on M3 Ultra
25.2
tok/sdecode
M3 Ultra (60-core GPU) + 96GB unified
rank in tier
26/30M3 Ultra runstop 87%
best workload
chat-short
where the rig flew
slowest workload
—
single-workload run
backend
llama.cppllama.cpp is the universal backend — broad model support, modest speed ceiling.
faster than
- llama3.2 on AMD64 Family 23 Model 104 Stepping 1, AuthenticAMD23.9 tok/s
- Muse-Glimmer-30B-ctx8k on RTX A400023.3 tok/s
- gemma-2-9b-it-4bit on M3 Pro22.0 tok/s