phi-4-4bit

on M3 Ultra

74.4

tok/sdecode

M3 Ultra (60-core GPU) + 96GB unified

rank in tier

11/20M3 Ultra runstop 55%

best workload

chat-short

where the rig flew

slowest workload

—

single-workload run

backend

mlxMLX is the fastest backend on Apple Silicon for dense models.

faster than

Qwen3.6-27B-Q4_K_M.gguf on RTX 509069.9 tok/s
claude-haiku-4.5 on M3 Pro45.1 tok/s
llama-3.3-70b-instruct on M3 Pro43.2 tok/s

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.