Chrome installs 4GB Gemini Nano

Google Chrome’s on-device AI rollout has become visible to users because the model is large enough to show up in storage. Reports published in early May said some desktop users found a roughly 4 GB `weights.bin` file inside Chrome directories tied to Gemini Nano, Google’s local language model. Google’s own developer documentation does not describe a user-facing prompt before that download. It says Chrome manages Gemini Nano downloads, updates and deletion automatically in the background. ### Where is the 4 GB file coming from? Google’s Chrome developer site says built-in AI in Chrome uses Gemini Nano, and that the browser “provides and manages” those models. The same documentation says the initial model download is triggered by the first call to a built-in AI API that depends on Gemini Nano, such as `Summarizer.create`. (malwarebytes.com) Chrome’s model-management page also says downloads can continue after a tab is closed, can resume after a browser restart within 30 days, and are updated automatically when Chrome starts. That means the model is treated more like a browser component than a conventional user-downloaded file. ### Is Chrome supposed to download Gemini Nano only when a site asks for it? (developer.chrome.com) Google’s documentation says yes, in the ordinary case: a site or app using a built-in AI API triggers the first download. But the same page adds an important exception. It says `availability` can sometimes trigger a download shortly after a fresh profile starts up if Chrome’s Gemini Nano-powered scam detection feature is active. (developer.chrome.com) Chromium security materials separately say the browser team added “an additional layer of protection” using the on-device Gemini Nano large language model on Chrome desktop to help detect tech-support scams. That ties the local model not just to web developer APIs, but to Chrome’s own security features. ### Which machines are eligible to get it? (developer.chrome.com) Google’s hardware requirements page says Gemini Nano-backed APIs work on desktop operating systems including Windows 10 and 11, macOS 13 or later, Linux, and some Chromebook Plus devices. The same page says users need at least 22 GB of free space on the volume containing the Chrome profile. Google expanded support in October 2025 by adding CPU inference in Chrome 140 on Linux, macOS and Windows, after initially focusing more heavily on GPUs. (chromium.org) That widened the pool of machines that could run Gemini Nano locally, even if performance differs by hardware. ### What has Google said publicly? Parisa Tabriz, Google’s vice president and general manager for Chrome, said in a public response reported by multiple outlets that Gemini Nano in Chrome powers on-device scam detection and developer APIs. (developer.chrome.com) Those reports said she described the model as local and framed it as part of Chrome’s security and developer strategy. Reuters could not independently verify the full text of the social-media thread from the source page available here, but the description is consistent with Google’s published Chrome documentation on local execution and scam detection. (developer.chrome.com) Google’s public Chrome AI pages also say no data is sent to Google or third parties when the local model is used. That claim appears in the hardware and setup documentation for built-in AI APIs. ### Why is this surfacing now? May 19 is the next obvious checkpoint. Google said on February 17 that Google I/O 2026 will run May 19-20 at Shoreline Amphitheatre in Mountain View, California, and that attendees should expect updates “from Gemini to Android and more.” (makeuseof.com) Chrome’s AI pages already position Gemini Nano as a browser-managed model for summarizing, writing, rewriting and prompting inside web apps. (developer.chrome.com) If Google adds more user-facing Gemini features in Chrome or elsewhere next week, the model-management system now drawing scrutiny is likely to get more attention, not less. (developer.chrome.com) (blog.google)