Chrome Installs 4GB Gemini Nano

About 4GB of OptGuideOnDeviceModel/weights.bin was downloaded to a Chrome user profile, and it was identified as the on-device model for Gemini Nano.

By cross-referencing macOS fseventsd, Chrome's Local State, ChromeFeatureState, and GoogleUpdater logs, evidence was presented that the background installation occurred without any user input. The Local State also recorded performance_class: 6 and vram_mb: 36864, leaving traces of a device suitability check.

• April 24, 2026, 16:38: OptGuideOnDeviceModel directory created
• April 24, 2026, 16:47: The unpacker wrote out weights.bin, manifest.json, verified_contents.json, and on_device_model_execution_config.pb
• April 24, 2026, 16:53: Moved to the final path OptGuideOnDeviceModel/2025.8.8.1141/weights.bin, with total installation time of 14 minutes 28 seconds
• It was reported that repeated behavior was observed where Chrome would re-download the model after deletion
• GoogleUpdater received a 7MB control component over HTTP, while it was explained that Chrome's in-process installer directly downloaded the actual weights

There were also traces of the OnDeviceModelBackgroundDownload and ShowOnDeviceAiSettings flags being enabled together, and it was pointed out that model deployment could begin before the settings UI is even exposed. Chrome's approach to on-device AI deployment is fueling growing controversy over storage, transparency, and privacy.