Practical upshot for 16 GB Macs: • Speed + reliability: Llama-3.1-8B (4-bit) • Best quality that fits: Gemma-3-12B QAT 3-bit — 5.9 GB, clean outputs • Skip: any 14B at naïve 3-bit — slower and less…

Three traps that almost gave me wrong numbers: • cold-start (warm up before timing) • the laptop sleeping mid-run (a "460s load" was a nap — use caffeinate) • memory accumulating across models in one…