The Unplug Kit — a Private ChatGPT on Your Mac in 10 Minutes: the No-Terminal Setup Card, the Two Privacy Switches, the Model for YOUR RAM, and 10 Starter Prompts
The companion to the video. Local AI finally installs like a normal Mac app — this kit is the whole path on printable pages. PAGE 2 — THE SETUP CARD: every click from ollama.com to your first answer (download the 175MB app, drag the llama into Applications, pick gemma4:12b in the 'Find model' box, send one message; the model downloads once, about 7.6GB, then lives on your disk). What you will NOT see: no terminal, no installer wizard, no account, no credit card. PAGE 3 — THE TWO SWITCHES that turn private from a feeling into a fact: 'Enable cloud models and web search' OFF (local models never ask you to sign in — only cloud features do), Context length raised for long documents, plus where your model files actually live (one folder you can open, move, or delete) — and THE RITUAL: Wi-Fi off, ask the question you'd never type into a website, watch it answer with the internet visibly dead. PAGE 4 — WHICH MODEL FOR YOUR MAC, from our real 5-test battery (2026-08-11, temp 0, transcripts kept): 8GB = llama3.2:3b for drafting only (it blessed a contract trap and invented a lab problem — receipts in the free Sealed Files); 16GB+ = gemma4:12b, the perfect 22/22 score at 27 tok/s in under 8GB; why bigger is just slower (the 15GB Mistral tied it at half the speed); and the mistake that freezes Macs — never pull a model bigger than your memory (our real 21GB-on-24GB freeze, Aug 10). PAGE 5 — 10 STARTER PROMPTS copy-paste: the contract sweep, the lab translator, the money order, the hard email, the honest summary, and five more. PAGE 6 — WHEN LOCAL IS NOT ENOUGH: the honest local/cloud/human rule. PAGE 7 — the 5 usual suspects and fixes. App flow verified on Ollama for macOS 0.32.9 (Aug 2026, macOS 14+). Independent kit from Hyperautomation Labs.
Free. No spam. Unsubscribe anytime.