llamaperf

Qwen3.8

on M5 Ultra 256GB · oMLX

Tone: negativeApple hardware
Nov 20, 2026
Throughput
5.5 t/s gen
Quant
oQ6e (MLX)
System RAM
256 GB

Summary

User reports Qwen3.8 at about 5.5 tokens/sec on a Mac Studio M5 Ultra with 256 GB unified memory when the GPU clock stays at 338 MHz during MLX inference. Setup is oMLX 0.7.3 with MLX 0.32.4, oQ6e weights and MTP speculation; GPU power sat at roughly 1.5-7.4 W in the degraded state. Restarting oMLX did not help; a full reboot restored the GPU to 1620 MHz and 60-70 W. The user says previous normal Qwen3.8 oQ6e-MTP runs on this machine were about 120-143 tokens/sec, and asks whether FileVault, SSH or a LaunchDaemon is involved.