AI on PAI

PAI ships a complete local AI stack. You can chat with an LLM, run models from the terminal, or hit the API from your own scripts — all without a network connection.

The short version

1. Launch Open WebUI from the app menu.
2. Pick a model from the dropdown. The first run downloads weights (internet needed here) and stores them on persistent disk.
3. Chat.

That's it. No account, no API key, nothing leaves the machine.

The pieces

• Ollama — the inference daemon. Bound to 127.0.0.1:11434, so it never accepts network traffic from other hosts.
• Open WebUI — a familiar chat interface running in your local browser.
• pai-recommend-model — a helper that picks a model that fits your RAM and GPU.

Deeper guides

• Choosing a model — match the model size to your hardware.
• Managing models — pull, delete, and pin versions.
• Using Ollama — CLI, HTTP API, Python SDK.
• Using Open WebUI — the chat UI in depth.

How privacy works here

Ollama listens on loopback only. The firewall (see ../architecture/network.md) drops inbound 11434 either way. No prompts, completions, or model metadata leave the machine. See ../architecture/ai-stack.md.

Offline considerations

Model weights are persistent — once you pull a model with network, it runs fine air-gapped. To pre-stage models for offline work, see ../privacy/offline-mode.md.