Chrome 4GB Gemini Nano on-device AI model is stored as a 4GB file named weights.bin in the OptGuideOnDeviceModel folder within Chrome user data. Deleting the file triggers Chrome to re-download it on the next restart, the download took roughly 15 minutes in an audit, and the behavior has been confirmed on three platforms: Windows 11, Apple Silicon Macs, and Ubuntu.
Deleting the 4GB Gemini Nano on-device AI model file causes Chrome to re-download the model automatically on the next browser restart with no notification or prompt. The model is stored as weights.bin in the OptGuideOnDeviceModel folder within Chrome user data. An audit reported that the download process took roughly 15 minutes. Deleting the file or folder does not provide a permanent solution because Chrome restores the model on restart.
This automatic restoration behavior has been confirmed on Windows 11, Apple Silicon Macs, and Ubuntu. On Windows, the file path is %LOCALAPPDATA%\Google\Chrome\User Data\OptGuideOnDeviceModel\weights.bin; on Mac and Linux it is in the equivalent Chrome profile directory. Users who do not want the model restored must disable the feature rather than delete the file. Chrome provides multiple disabling options.
Available methods to prevent the automatic re-download include disabling the feature via chrome://flags, turning off the On-device AI toggle in Settings > System, or setting the Windows registry key OptimizationGuideModelDownloading to disabled. Unless one of these measures is used, Chrome will restore the model after deletion.
Chrome’s local Gemini Nano model powers several on-device AI features in Chrome. The model is used for “Help me write an email”, scam detection, smart paste, page summarization, and AI-assisted tab grouping. These features run on the device rather than routing queries to external servers when they use the local model. The on-device model is stored as a file named weights.bin in the OptGuideOnDeviceModel folder within Chrome user data. The local model provides functionality that operates within the browser on users’ machines.
Chrome’s AI Mode pill in the address bar routes every query to Google’s cloud servers, and the local Gemini Nano model does not power AI Mode. Users are paying the storage and bandwidth cost for a feature they’re not actually using privately. The local Gemini Nano model and AI Mode are distinct operational paths within Chrome.
Alexander Hanff raised privacy concerns about Google storing the on-device Gemini Nano model on users’ machines. He argues Google may be violating Article 5(3) of the ePrivacy Directive and Articles 5(1) and 25 of the GDPR by storing data on the device. Hanff referenced the requirement for prior, freely-given, specific, informed, and unambiguous consent in his critique.
Hanff also linked a separate case involving Anthropic’s Claude Desktop, which he described as involving silent pre-authorization of browser automation. He said the Claude Desktop case affected roughly three million user machines. Hanff presented the Anthropic comparison alongside his concerns about Chrome’s behavior.
These statements summarize the privacy concerns and legal implications raised by Alexander Hanff. They reference specific EU legal provisions and a comparable Anthropic case.
