The main promise of local AI is that your prompts and files are processed on your own computer and never sent to a company’s servers. That is true of the model itself. Whether it is true of the whole setup depends on the apps around it.
What stays on your machine
- Your prompts, the model’s answers and any documents you give it, while the model runs locally.
- Your conversation history, which apps store in files on your disk. Delete it there if you want it gone.
What can still touch the network
- Downloading models. Fetching a file from Hugging Face or an app’s catalog is a network request, though it does not include your prompts.
- Updates and telemetry. Many apps check for updates and some send usage data. Look in the settings for options to turn them off, and read each app’s privacy policy.
- Tools and agents. A model connected to web search, a browser or an online API will send queries out. A coding agent may also be configured to use a hosted model. Check which backend an app is using.
- Servers. A local API server (guide) listens on your machine by default. Do not expose it to the network without authentication.
Running fully offline
- Download the app and the model files while you are online. Keep the installer.
- Turn off automatic updates and usage reporting in the app.
- Disconnect, or block the app in your firewall, and confirm that chat still works.
- Do not enable web search or other online tools for that setup.
File safety
Prefer GGUF and safetensors files: they store data only. Older PyTorch .bin or pickle files can contain code that runs when loaded, so only open those from sources you trust. Download from the model’s official page, and check that the file size and checksum match when they are published (formats).
For a sense of what each app supports, see the apps page.