Get the Local AI Cheat Sheet

Running AI locally isn't about your processor — it's about MEMORY, and three numbers decide whether it feels like magic or misery. This one-page field guide is the whole thing, beat for beat with the video: (1) the memory ceiling — a rule of thumb of ~0.6 GB of RAM per billion parameters, so a 7B model needs ~4-5 GB and a 14B needs ~8-10 GB, minus what your OS already uses; (2) quantization — the near-free 'Q4' trick that shrinks a model ~70% (a 7B drops from ~14 GB at full precision to under 5 GB) with barely any quality loss, handled for you by default in Ollama; (3) the context window — the silent killer that also lives in RAM on top of the model, so a smaller model with headroom beats a bigger one that's choking. Then the actionable part: a pick-your-model-by-RAM table (8 GB → 3-4B, 16 GB → 7-8B like Qwen2.5-Coder, 24-32 GB → 14B, 48 GB+ → 30B+), the exact three-step setup (download Ollama, run 'ollama run qwen2.5-coder:7b', start typing — same on Mac, Windows and Linux), a handy-commands reference, and four free power upgrades (plug it into VS Code with Continue or Cline, use MLX builds on Apple Silicon, turn on KV cache quantization for longer chats, and go one size down when in doubt). Measured live on an Apple Silicon laptop: Qwen2.5-Coder 7B, 4.7 GB on disk, ~48 tokens/sec, 100% on-device, $0.

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — one practical AI email per week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.