Since the beginning of 2025, Google has been steadily deploying the Gemini Nano model to Chrome browser users. This sophisticated model operates locally on the device, primarily serving to identify fraudulent or phishing websites while powering various AI-driven features. However, Google conspicuously failed to notify users in advance. Instead, Chrome silently downloaded the model weight files in the background. Remarkably, these files consume approximately four to five gigabytes of storage per installation.
For instance, when we installed the Stable, Beta, and Canary versions of Chrome simultaneously on macOS, each iteration autonomously downloaded its own dedicated set of model weight files. Consequently, the actual storage space consumed exceeded a staggering fifteen gigabytes. Furthermore, manually deleting these files proved entirely futile, as Chrome would persistently redownload them automatically. Fortunately, Google now provides a definitive toggle switch to disable on-device AI, thereby halting these automatic downloads.
New Storage Prerequisites for AI Models
Recently, Google updated the official Help Center documentation concerning Chrome’s on-device generative AI models. As reported by Windows Report, the revised documentation stipulates that installing and utilizing models supporting local AI functionalities now necessitates approximately 20GB of available disk space. Additionally, this process requires an unmetered data connection and adequate device performance.
While Google remains somewhat ambiguous regarding the precise definition of “adequate device performance,” they emphasize a crucial point. Chrome will only initiate the automatic download of the requisite model weight files if the host machine meets these stringent performance criteria. It is important to clarify that the model weight files themselves remain approximately four to five gigabytes in size. However, Chrome will only trigger the download process if the available disk space comfortably exceeds the 20GB threshold.
Disabling the Feature is the Only Solution
Google explicitly emphasizes within their official support documentation that users should refrain from manually attempting to delete the model weight files downloaded by Chrome. Deleting them merely prompts Chrome to automatically attempt a fresh download. To completely disable the on-device AI functionality and permanently prevent the downloading of these files, users must navigate to the Chrome browser settings.
From there, proceed to System, and then select On-Device AI. Once the on-device AI feature is officially toggled off, the system will automatically purge any previously downloaded model weight files. This is the only effective method to reclaim your storage space and prevent future unauthorized downloads.
Verifying the Successful Deletion of Weight Files
According to comprehensive testing conducted by our team, Google instantly and automatically obliterates the locally downloaded model files immediately after the on-device AI feature is disabled. Therefore, users can proactively verify the removal by navigating to the designated model weight folder after disabling the function. Under normal circumstances, the entire folder will be completely eradicated.
If you manually navigate to the OptGuideOnDeviceModel directory and find it empty, this definitively confirms that disabling the on-device AI has taken effect. At this juncture, Google has already automatically deleted the corresponding version folder, which originally housed the massive model files. For reference, you can check these specific file paths:
For Windows 10/11 users, check the following path: C:\Users\{Your Username}\AppData\Local\Google\Chrome\User Data\OptGuideOnDeviceModel\2025.8.21.1028\weights.bin
For macOS users, check these two paths depending on your installation: /Users/{Your Username}/Library/Application Support/Google/Chrome/OptGuideOnDeviceModel/2025.8.8.1141/weights.bin or /Users/{Your Username}/Library/Application Support/Google/Chrome Canary/OptGuideOnDeviceModel/2025.8.8.1141/weights.bin
Support Our Threat Intelligence
If you find our CVE report and cybersecurity news helpful, consider supporting our work.