M5 Max LLM benchmark
On the M5 Max, measured single-user chat decode speeds are: ~35.5B models at 88.3 tok/s (Q6_K, ollama; Measured, 1 result, source run). Short-chat workload, local runtimes, dense models only; quantization, backend and settings vary between submissions, so check each row in the tables below.
Fastest single result in these submissions (any model size)
88.3 decode tok/s
qwen3.6-agent-q6k-ctx128k (35.5B) via ollama (Q6_K). Workload: chat-short. Reported hardware: M5 Max (40-core GPU) + 48GB unified. see full run
qwen3.6-agent-q6k-ctx128k
| Workload | Model size | Reported hardware | Backend | Quant | decode tok/s | prefill tok/s | TTFT | Run |
|---|---|---|---|---|---|---|---|---|
| chat-short | 35.5B | M5 Max (40-core GPU) + 48GB unified | ollama@0.32.1 | Q6_K | 88.29tok/s | 16.20tok/s | 6,915ms | r_udwgba7udqu |