TL;DR

• Chrome's Prompt API downloads Gemini Nano model files up to 4 GB to local storage without a visible prompt.
• The download happens in the background on eligible hardware, even if you never use any AI feature.
• You can disable it via a specific chrome://flags toggle or a group policy entry.

Chrome is silently downloading 4 GB of local AI model files

What's happening

Google Chrome has been quietly downloading Gemini Nano, its on-device large language model, onto users' machines. The files total roughly 4 GB and land in Chrome's profile directory with no download notification in the browser UI.

The behavior was documented by OS X Daily in May 2026, drawing on user reports of unexpected disk consumption. Chrome's built-in AI features, collectively exposed through the Prompt API, use Gemini Nano to power summarization, writing assistance, and similar tasks directly in the browser. When Chrome determines the device meets the hardware requirements (typically a machine with at least 22 GB of storage free and a capable GPU), it schedules the model download automatically.

The download runs through Chrome's component updater system, the same mechanism that pushes CRLSets and other background updates. Because component updates don't surface in chrome://downloads, most users don't notice until they audit disk usage.

Why it matters for extension developers

Extensions that use the built-in AI APIs, specifically window.ai and the Prompt API, depend on Gemini Nano being present on the device. If a user has disabled the download or is on ineligible hardware, calls to ai.languageModel.create() return an "unsupported" availability status rather than a ready model.

That's a real compatibility surface. An extension built around Chrome's on-device inference will silently degrade for a non-trivial slice of users: those who blocked the download, those on low-storage machines, and those on Chrome for Linux (where Gemini Nano support is limited as of mid-2026).

The 4 GB footprint also creates friction for users on metered connections or smaller SSDs. Complaints in the OS X Daily report and on Hacker News threads center on the lack of consent, not necessarily the feature itself. If your extension relies on the Prompt API, that user sentiment is worth factoring into your support queue planning.

How to disable the download

OS X Daily's guide covers three approaches, which align with what's documented in Chrome's own flags interface.

Via chrome://flags: Navigate to chrome://flags/#optimization-guide-on-device-model and set it to "Disabled." Chrome will stop downloading the model and won't use any cached portions already on disk. Restart Chrome to apply.

Via chrome://components: Open chrome://components, find "Optimization Guide On Device Model," and select "Check for update" to inspect its current state. There's no direct delete button here, but disabling the flag above prevents future fetches.

Via enterprise policy: Admins managing Chrome through group policy or chrome.policy can set OptimizationGuideModelDownloadingEnabled to false. This is the reliable path for managed fleets and for developers who want to keep their own machines clean during extension testing.

The model files themselves sit inside the Chrome profile at a path like ~/Library/Application Support/Google/Chrome/OptimizationGuide on macOS or %LOCALAPPDATA%\Google\Chrome\User Data\OptimizationGuide on Windows. Deleting the folder manually reclaims the disk space, though Chrome will re-download it unless the flag is disabled first.

The broader pattern: Chrome as an AI runtime

This isn't an isolated decision. Google has been positioning Chrome as an inference runtime since the Chrome 122 era, with the Prompt API, Summarization API, Translation API, and Writing API all moving through origin trials toward stable. Gemini Nano is the shared model underneath all of them.

The component-updater delivery mechanism was chosen specifically because it bypasses the Chrome Web Store review pipeline and doesn't require a user-visible install action. That's efficient for Google, but it means the 4 GB lands on disk before most users have any reason to know the feature exists.

For extension developers, the practical upshot is this: don't assume Gemini Nano is present. Always check availability before calling the Prompt API, and build a fallback path for the "unsupported" and "downloadable" states. The "downloadable" state means the model hasn't been fetched yet, and triggering ai.languageModel.create() in that state will kick off the download mid-session, which is its own user experience problem.

Extensions that do AI inference entirely through their own backend avoid this dependency entirely. The tradeoff is latency and API cost versus the unpredictable availability of the on-device model.

The original OS X Daily article is at https://osxdaily.com/ and covers the macOS-specific file paths and UI steps in more detail.

For context on how AI-powered extensions are pricing and packaging their inference costs, see the related post on AI extension pricing patterns.