ChatGPT needs the internet to respond. This AI doesn't — and it just proved it.

The exact clicks

1. Install Ollama: on Mac or Linux, paste curl -fsSL https://ollama.com/install.sh | sh into a terminal and press Enter. On Windows, use PowerShell and paste irm https://ollama.com/install.ps1 | iex instead.
2. Once it finishes, download a model with a network connection: type ollama pull llama3.2 and press Enter. This step needs the internet — it's the only one that does.
3. Optional, to prove it to yourself: turn off WiFi (or unplug ethernet) once the download finishes.
4. Type ollama run llama3.2 and press Enter. It starts and answers with zero network connection.
5. Ask it anything — a question, a coding request, a word problem. The answers are real, generated entirely on your machine.
6. To try a different model instead: ollama pull phi3 (smaller, faster) or ollama pull gemma2 (a middle ground) — swap the model name in both the pull and run commands.

What changes

• This is standard, current Ollama behavior — verified live: v0.32.5 as of this episode, free and open-source (MIT license).
• Only ollama pull and ollama push need a network connection. Once a model is downloaded, ollama serve and every query after it run with no network at all — confirmed in this episode inside an isolated network namespace with a verified-failed outbound connection.
• The model itself has no way to know it's running locally — asked directly whether it's connected to the internet, it answered (wrongly) that it's a cloud-based model that needs one. This isn't a bug specific to one model; language models generally have no built-in way to introspect their own deployment.

What it will not do

• The one-time model download does need real internet — this is an offline-after-setup tool, not a zero-download one.
• A modern laptop with 16GB RAM comfortably runs 7-8B parameter models; larger models (Llama 4, some Qwen 3 variants) need considerably more.
• Model names move fast — this episode uses llama3.2, the current official quick-start model as of this episode; check ollama.com's model library for the current recommended default before assuming an older name (llama3, phi3) is still the best starting point.