YOUR OWN AI WORKSHOP
Put the M1
on the bench.
By Alvin
The 64GB M1 Max ran the same Qwen and Llama model files and prompts as the primary Mac. Ollama 0.33.2 loaded one model at a time. As on page 63, the scores count four test examples and elapsed time includes model loading.
| MODEL / BRIEF | CHECKS | ELAPSED s |
|---|---|---|
| Qwen3.5 4B / notebook | 4/4 | 3.93 |
| Qwen3.5 4B / triage | 4/4 | 3.45 |
| Llama3.2 3B / notebook | 4/4 | 1.67 |
| Llama3.2 3B / triage | 0/4 | 1.74 |
Qwen passed both sets of four examples. Llama passed the notebook examples, then returned triage groups as ID-based lookup tables instead of the requested lists of records. Even its empty group used the wrong format.
Qwen generated replies at about 65–66 tokens/s; Llama about 101. Other work and downloads were running on the Mac. The primary Mac also used a different Ollama release. These are observations of those runs; a hardware ranking needs matched software and controlled workloads.