[Official FAQ] Share your experience or follow-up questions below.

The main benefit of local inference is that your prompts and outputs are processed on your own machine and do not need to be sent to a third-party service. But "local" is not automatically "secure". Consider the following.

• App network behavior: some apps check for updates, download models, or include optional telemetry. Review the settings, and if it matters, watch connections with a firewall or network monitor.
• Exposed API: binding a server to 0.0.0.0 or forwarding a port can make it reachable by others. Local APIs often have no authentication. If you need remote access, use a reverse proxy with authentication and HTTPS, or a VPN.
• Plugins and tools: web search, cloud sync and extensions may send some data out. Check what each one does.
• Model sources: download from official or reputable publishers. Prefer formats like GGUF and safetensors over older pickle-based files. Be careful with repositories that require trust_remote_code, and read the code first.
• Stored history: chat logs are usually saved on disk. On shared or unencrypted machines, that is a leak risk.

For sensitive work (contracts, medical notes, source code): test once with the network disconnected to confirm everything works offline, keep the server bound to localhost, and clean up history regularly.

Local also does not make the model's answers more correct, so keep verifying important output.

What privacy practices do you follow? Add your tips below.