Chrome Quietly Put A 4GB AI Model On Your PC: Find And Kill It

Four gigabytes. That is what Google Chrome quietly slid onto roughly a billion desktops over the past year, with no prompt, no checkbox, no notification. The file is called weights.bin. It lives in a folder named OptGuideOnDeviceModel inside your Chrome user data directory. And until February 2026, deleting it did nothing. Chrome would just pull it down again on the next restart.

The file is Gemini Nano, Google’s smallest on-device language model. Privacy researcher Alexander Hanff’s forensic write-up of the silent Gemini Nano install on May 4, 2026 set off the firestorm, but the install itself has been running since 2024. What followed is the strangest part of the story. The most visible AI feature in Chrome, the AI Mode pill in the address bar, doesn’t use the local model at all. Those queries fly straight to Google’s cloud.

Here is what is actually on your machine, why it landed there, and the exact clicks that get rid of it in May 2026.

What Chrome Actually Put On Your Computer

The artifact is a single binary. On Windows it sits at %LOCALAPPDATA%\Google\Chrome\User Data\OptGuideOnDeviceModel\weights.bin. On macOS it lives in the parallel path inside the Chrome profile folder, and on Ubuntu it appears in the equivalent Linux profile directory. The reported weight on disk runs between 3 and 4 gigabytes depending on the build.

Hanff caught the download in an unusually clean way. He spun up a brand-new Chrome profile on an Apple Silicon Mac, opened a single tab, and walked away. Then he combed the macOS .fseventsd kernel log, which records every file event at the system level. The full 4 GB model assembled itself in roughly fifteen minutes with zero human input.

The fact-checking outfit Snopes ran the same check on six staff laptops. Three of the six found the file present, two on macOS and one on Windows, matching Hanff’s path exactly. The model is not on every Chrome install. It is on every install that meets the spec.

The Spec Sheet Almost Nobody Saw

Chrome silently runs a hardware audit before pulling the model. The Chrome for Developers built-in AI requirements page lists the bar in plain text. If your machine clears it, the download starts.

• Storage: at least 22 GB free on the volume holding the Chrome profile
• GPU: strictly more than 4 GB of VRAM, or a fallback path of 16 GB system RAM plus four CPU cores
• Operating system: Windows 10 or 11, macOS 13 or later, Linux, or Chromebook Plus
• Network: an unmetered connection for the initial pull

The 22 GB headroom requirement is the strangest line on the page. The model itself ships at roughly 4 GB, so Chrome is asking for more than five times the file’s footprint just to start. Independent developers digging into the docs have flagged the gap without a clean answer from Google. Mobile is excluded entirely. Chrome for Android, Chrome for iOS, and ChromeOS on standard Chromebooks do not run Gemini Nano.

The AI Mode Paradox

Here is the part that lands hardest with privacy lawyers. The AI feature most Chrome users actually see, the AI Mode button in the omnibox, never touches the local model.

Those queries route to Google’s cloud servers. The 4 GB on your laptop powers a different, narrower set of features: the Help me write composer in text fields, on-device scam detection on incoming pages, smart paste, page summarization, and the Prompt API exposed to web developers. Decrypt’s reporting on Chrome’s deleted privacy promise caught Google quietly rewriting the help-page line that used to say data “is not sent to Google’s servers.” The new wording acknowledges that websites calling the on-device API can see inputs and outputs governed by their own policies.

Why Hanff Calls This Illegal In Europe

The legal frame Hanff is leaning on is Article 5(3) of Directive 2002/58/EC, the ePrivacy Directive. That is the same clause behind every cookie banner you have ever clicked through. It bars storing or accessing information on a user’s terminal equipment without prior, freely given, specific, informed, and unambiguous consent, with one narrow exception for what is strictly necessary to deliver a service the user explicitly asked for.

His argument is short. The 4 GB model is information. It sits on your terminal equipment. You did not consent. Chrome works fine without it. The strict-necessity carve-out does not apply.

This is, in my professional opinion, a direct breach of Article 5(3) of Directive 2002/58/EC, a breach of the Article 5(1) GDPR principles of lawfulness, fairness, and transparency, a breach of Article 25 GDPR’s data-protection-by-design obligation, and an environmental harm of a magnitude that would be a notifiable event under the Corporate Sustainability Reporting Directive.

Hanff told Snopes he intends to file criminal charges in his home jurisdiction. He has already issued a similar cease-and-desist to Anthropic over the Claude Desktop browser bridge, suggesting he sees a pattern across AI vendors rather than a single bad Google decision.

The potential exposure is not small. GDPR caps fines at the higher of 20 million euros or 4 percent of global annual revenue. For Alphabet that ceiling sits north of 12 billion dollars. No regulator has opened a formal probe yet, and the analysis is one privacy professional’s reading, not a court ruling. But Article 5(3) cases over silent device storage are exactly the cases EU regulators have been winning.

The environmental angle is the wildcard. Hanff estimated emissions of 6,000 to 60,000 tonnes of CO2 equivalent across a deployment in the hundreds of millions, a range wide enough to invite peer review but striking even at the low end.

Google’s Answer, Translated

Parisa Tabriz, Chrome VP and general manager, posted the company line on May 7. The on-device model, she said, exists to keep sensitive features like scam detection running without sending page contents to Google. The model frees itself up if storage runs low, and a settings switch to turn it off has been rolling out since February.

The full statement Google provided to Snopes ran on a similar track. It does not address the consent question Hanff raised at all. The wording defends what the model does. It does not defend how the model arrived.

How To Check And Remove It Right Now

Three minutes of clicking gets you to a clean answer. Open Chrome and paste chrome://on-device-internals into the address bar.

1. Enable debugging. If the page warns that internal debugging is off, click the link it provides and tick the Enable internal debugging pages box.
2. Open the Model Status tab. If the Foundational Model State reads No On-Device Feature Used, the model is not on your machine. You are done.
3. If it shows installed, open Chrome’s main Settings menu, click the System tab, and look for an On-System AI or On-device AI toggle. Switch it off.
4. The toggle deletes the file. Snopes staffers who flipped the switch saw the OptGuideOnDeviceModel folder vanish immediately, with no reinstall observed in the days that followed.
5. If the toggle is missing, the rollout has not reached you yet. Open chrome://flags, search for optimization-guide-on-device-model, and disable it along with the Prompt API, Summarizer API, Writer API, Rewriter API, and Proofreader API flags before relaunching.

On managed Windows machines, IT teams can push the policy through a registry edit setting OptimizationGuideModelDownloading to disabled. The Google Chrome community thread on the OptGuideOnDeviceModel folder has been collecting user reports and workarounds since the file first started showing up.

One catch worth flagging. Tom’s Hardware reporters testing on Chrome v147 couldn’t see the toggle on a MacBook running the same version where it appeared on a Windows machine. The February rollout is gradual. If your Mac is missing the switch, the flags route is the only working path today. Check back after the next stable Chrome release before assuming you are stuck.

This story sits inside a wider pattern around silent AI deployment, the same fault line we covered when Utah’s SB73 VPN crackdown reshaped consumer privacy defaults earlier this month, and again when Indian regulators flagged on-device AI behavior in SEBI’s Claude Mythos cyber-suraksha task force. The browser is not a neutral pipe anymore. It has opinions about what should run on your hardware, and increasingly, it makes those decisions before asking.

Frequently Asked Questions

Is the 4 GB Gemini Nano file dangerous?

No. The file itself is not malware and does not exfiltrate data on its own. Inputs to the on-device APIs stay local unless a website chooses to call them, in which case that site’s privacy policy governs the data. The complaint is consent and disk usage, not malicious behavior. If you want it gone for principle or storage reasons, use the toggle in Settings, System, On-device AI, or disable the optimization-guide-on-device-model flag at chrome://flags.

Will Chrome redownload the file after I delete it?

It depends on how you removed it. Manually deleting the OptGuideOnDeviceModel folder without disabling the underlying setting triggers a fresh download on the next Chrome restart. Using the new February 2026 toggle in Settings, System, On-device AI stops the redownload according to Google and matches the behavior Snopes observed across multiple test machines. Confirm by reopening chrome://on-device-internals a day or two later and checking that Foundational Model State still reads No On-Device Feature Used.

Does this affect Chrome on my phone?

No. Gemini Nano in Chrome is desktop-only. 9to5Google’s breakdown of the 4GB Chrome AI storage rules confirms that Chrome for Android, Chrome for iOS, and standard ChromeOS Chromebooks are not part of the deployment. Only Windows 10 and 11, macOS 13 and later, Linux, and Chromebook Plus devices that clear the hardware bar receive the model. If you only use Chrome on a phone, you have nothing to remove.

Why does Chrome ask for 22 GB free when the file is only 4 GB?

Google has not given a public reason. Independent developers reading the requirements doc have flagged the gap as unexplained. The plausible answers are temporary build files during model assembly, future expansion headroom for larger Nano variants, or a defensive buffer to keep the OS itself stable on near-full disks. Until Google clarifies, treat 22 GB free as the trigger threshold. Drop below it and Chrome will refuse to install the model, or auto-delete an existing one.

Can my employer or school force the model onto a managed Chrome?

Yes, and they can also force it off. Chrome enterprise policy includes an OptimizationGuideModelDownloading registry key on Windows and equivalent plist controls on macOS. IT administrators can pin the model on, off, or leave it user-controlled. If you run a personal Chrome profile on a managed device, the policy your IT team sets overrides the in-browser toggle. Ask your admin if you cannot find the switch and the flags workaround also seems blocked.

The bigger lesson is the one the toggle does not solve. A browser that ships hardware-tier AI to a billion machines without a checkbox is making a product decision that used to require a conversation. Google may yet win the legal argument in Europe. The trust argument is harder, and Chrome will be defending it again the next time a 4 GB file lands on a desktop somewhere with no one’s permission but its own.