Chrome downloads 4GB Gemini Nano model silently

Multiple technology outlets report that Google Chrome has been downloading a 4GB on-device AI model named Gemini Nano into users' local Chrome data directories, typically as a file called weights.bin inside an OptGuideOnDeviceModel folder, and doing so without an obvious consent prompt (The Verge; CNET; PCMag). Security researcher Alexander Hanff flagged the rollout and timelines in public posts, and reporting places the apparent surge in appearances between late April and early May 2026 (CNET; The Verge). Ars Technica notes that Chrome's local AI features, and local model installs, have been present in some form since 2024, so some machines may already have had local models for years (Ars Technica).

The reported file is weights.bin, about 4GB in size, and sits under Chrome's user data directories in an OptGuideOnDeviceModel folder on affected machines, according to multiple reports and user troubleshooting threads (The Verge; PCMag; Google support thread). Chrome appears to download the model when certain on-device AI features are enabled and when the device meets hardware conditions cited in reporting; outlets cite Chrome flags and hardware/account signals as contributing factors (Ars Technica; The Verge). A Google spokesperson told CNET, "In February, we began rolling out the ability for users to easily turn off and remove the model directly in Chrome settings," and added that the model "will automatically uninstall if the device is low on resources" in the spokesperson's quoted remarks (CNET; PCMag).

Reporting and community troubleshooting recommend first disabling on-device AI features in Chrome via Settings > System > On-device AI or by visiting chrome://settings/ai and turning related toggles off, which reporting says prevents further downloads (The Verge; CNET; Google support thread). Users who find the file may manually delete the OptGuideOnDeviceModel folder or uninstall Chrome; however, outlets warn that Chrome may re-download the weights if on-device AI remains enabled or if flags trigger the behavior (The Verge; PCMag; Google support thread). The Google community support thread and multiple guides also point to chrome://flags searches for on-device or optimization flags as an additional step to stop automated installs (Google support thread; XDA; MacRumors).

Local, on-device models like Gemini Nano trade cloud compute and latency for local storage and CPU/RAM usage. Industry observers note that shipping smaller, local transformer-based models to endpoints reduces network round trips and can improve privacy for some flows, but it also creates measurable storage and energy footprints on end-user devices. For practitioners, this pattern reinforces an operational split: on-device inference reduces cloud costs and latency, while increasing client-side deployment complexity, update churn, and user-visible resource consumption.

Industry reporting frames the story as part technical nuisance and part privacy/regulatory concern: researchers and privacy advocates quoted in coverage argue that silent installs without clear consent may run into legal or trust issues in jurisdictions with strict data and consumer protections (CNET; Tom's Hardware). The episode highlights the UX tension between background feature enablement and transparent user consent for sizeable local downloads.

Observers should track whether Google adjusts Chrome's UI or opt-in flow to make local model downloads and sizes more visible, whether the company publishes clearer documentation about hardware eligibility conditions (Ars Technica; The Verge), and whether regulators or privacy groups escalate complaints in the EU or other jurisdictions (CNET; Tom's Hardware). For administrators and power users, monitoring Chrome release notes and chrome://flags changes will be the practical signals that the download behavior or controls have been updated.