Hermes 3 vs Gemma 4 vs Qwen3: I A/B tested 6 tasks on a 16GB Mac mini
Five A/B tests on real data, one month into running local AI. There is no single "agent model" — match work to model.
minillm.dev
Local AI on a 16GB Mac mini. One month in, still figuring it out.
I bought a $599 Mac mini on April 23, 2026, and started pulling models with Ollama. Now I run 13 daemons on it — job application grading, hobby research, daily briefings, Korean phrase drills. Used Claude as a dev partner the whole way. No NVIDIA cards, no Studio, no datacenter.
This is where I write down what worked, what didn't, and what I wish I knew on day one. If you're starting your own local AI thing, you might find something useful here.
Want the whole setup as a Notion template? The 16GB Mac Mini Local AI Setup — pay what you want on Gumroad ($3 minimum). Living template, free updates.