On-device AI

Instant on-device AI uses the model built into Chrome (Gemini Nano) to answer tiny requests locally, with no network call and no credits. Today it powers the suggestions you get when you type gpt and a question in the address bar: after three characters, up to three completions are proposed by the local model.

1. Open Settings from the side panel.
2. Switch Instant on-device AI on.
3. The first time, Chrome downloads the model. The setting shows Downloading model NN% until it is ready.

The toggle is only shown when your Chrome can run the model. Requirements set by Chrome: a recent Chrome (138 or newer), about 22 GB of free disk space for the download of a model of 2 to 4 GB, and a capable GPU or CPU. On computers that do not qualify, the setting is hidden and everything keeps working through the cloud.

Requests handled on-device never leave your computer. Everything else in GPT Workspace, including the chat, the AI bars and the Gmail features, runs on GPT Workspace’s servers and your Google account, as described in Data and privacy.