Google's Chrome browser has long been known for its appetite for storage, but a new discovery reveals it is quietly consuming even more space. A local AI model, identified as weights.bin, is silently downloaded onto both Windows and Mac systems upon installing Chrome, taking up approximately 4GB of storage. This file is part of the Gemini Nano on-device AI system, designed to perform tasks directly on the user's machine.
What is the weights.bin file?
The weights.bin file contains the neural network weights for Gemini Nano, Google's local AI model. Unlike cloud-based versions of Gemini, this models runs entirely on the user's device, enabling faster processing and potentially greater privacy for certain tasks. On macOS, the file resides at ~/Library/Application Support/Google/Chrome/OptGuideOnDeviceModel/. Windows users can find it at C:\Users\\AppData\Local\Google\Chrome\User Data\OptGuideOnDeviceModel\. In a test on a Mac, the file occupied 4.27GB of space.
How to remove the 4GB AI model
Simply deleting the weights.bin file is ineffective because Chrome will automatically reinstall it at the next opportunity. However, users can permanently remove the model by changing a single setting. Navigate to Chrome's Settings > System and toggle the “On-device AI” option to Off. In our test, this immediately erased the weights.bin file. The downside is that disabling this feature also turns off Chrome's local AI capabilities, including text suggestions and scam warnings.
What does Gemini Nano do?
According to a Google support page, Gemini Nano is responsible for several browser enhancements: summarizing web pages, organizing tabs, warning users about online scams, and offering writing assistance or text rephrasing as they type. Running AI locally offers lower latency and potentially greater privacy because data does not leave the device. However, Chrome may still share some browsing activity with Google's servers.
How does it compare to other AI models?
Gemini Nano's 4GB footprint is relatively modest compared to other local models. For instance, the 31 billion-parameter version of Google's Gemma 4 consumes about 20GB of storage, while the 128 billion-parameter Mistral Medium takes up a massive 80GB. This highlights that even smaller local AI models can be storage-intensive.
Privacy concerns and the bigger picture
The silent download of the weights.bin file has sparked debate. The fact that Chrome installs such a large file without asking for user consent has been criticized, especially given the privacy implications. While on-device AI can reduce reliance on cloud servers, it also introduces new privacy considerations about what data the browser collects and shares. Moving forward, users can expect more desktop applications to download local AI models, making it important to understand how to manage these files and their impact on system resources.
Source: PCWorld News