M1 Pro LLM benchmark
On the M1 Pro, measured single-user chat decode speeds are: ~1.2B models at 103.1 tok/s (Q8_0, 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)
103.1 decode tok/s
llama3.2 (1.2B) via ollama (Q8_0). Workload: chat-short. Reported hardware: M1 Pro (16-core GPU) + 16GB unified. see full run
llama3.2
| Workload | Model size | Reported hardware | Backend | Quant | decode tok/s | prefill tok/s | TTFT | Run |
|---|---|---|---|---|---|---|---|---|
| chat-short | 1.2B | M1 Pro (16-core GPU) + 16GB unified | ollama@0.32.14 | Q8_0 | 103.1tok/s | 8.77tok/s | 14,247ms | r_w-p7v--eync |
| chat-long | 1.2B | M1 Pro (16-core GPU) + 16GB unified | ollama@0.32.14 | Q8_0 | 81.35tok/s | 1,246.5tok/s | 2,529ms | r_w-p7v--eync |
| concurrent-decode | 1.2B | M1 Pro (16-core GPU) + 16GB unified | ollama@0.32.14 | Q8_0 | 97.26tok/s | no data | no data | r_w-p7v--eync |
| agent-trace | 1.2B | M1 Pro (16-core GPU) + 16GB unified | ollama@0.32.14 | Q8_0 | 93.14tok/s | 2,975.3tok/s | 516ms | r_w-p7v--eync |