Chrome’s Silent 4GB AI Install: How to Find It & Use It

Hot take: the part everyone’s outraged about – the 4GB silent install – is the least interesting story here. Google quietly handed you a free, local LLM your apps can call without an API key, and most coverage is busy telling you how to delete it. Both options are valid. You should know how to do both.

This story blew up in early May 2026 after privacy researcher Alexander Hanff documented the behavior. His original post kicked off coverage across Malwarebytes, TechSpot, 9to5Google, and Hacker News. Google’s response was essentially: this has been happening since 2024.

The 60-second context

Chrome has been writing a file called weights.bin – about 4GB of Gemini Nano model weights – into your profile under a folder named OptGuideOnDeviceModel. Malwarebytes confirmed the same pattern on Windows, Apple Silicon, and Ubuntu. No prompt, no notification. Hanff’s filesystem audit clocked the install at 14 minutes 28 seconds on a fresh profile that received zero human input.

The model powers on-device features – “Help me write,” scam detection, smart paste, page summaries, tab grouping. The kicker most articles bury: the prominent “AI Mode” pill in the omnibox doesn’t use the local model at all. It routes to Google’s cloud. So you’re paying the storage cost for features you may never trigger, while the visible AI button you do click sends data to servers anyway.

Step 1: Confirm it’s on your machine

Before you do anything, check whether you’re actually affected. Not every device qualifies – hardware requirements are strict (more on that below), which is why some users won’t find the file at all.

Open a Chrome tab and paste this:

chrome://on-device-internals

This internal page shows the file path, the current model version (something like 2025.8.8.1141), and an Uninstall button. If the page is locked on your build, open chrome://chrome-urls/ first to enable internal pages.

To find the file directly:

• Windows: %LOCALAPPDATA%GoogleChromeUser DataOptGuideOnDeviceModel
• macOS: ~/Library/Application Support/Google/Chrome/OptGuideOnDeviceModel (confirmed on Apple Silicon; path may vary on Intel Macs)
• Linux: equivalent path inside your Chrome profile directory

Step 2: Decide – kill it or use it?

Two paths. Pick one based on whether you want the disk space back or whether you’d rather get something out of the bytes you’ve already paid for.

Path A – Remove it (and keep it gone)

Just deleting weights.bin doesn’t work. Chrome’s Component Updater treats it as a managed file and pulls it back on the next restart or AI feature trigger. Worse: the file sometimes ships with a read-only attribute set, so on macOS/Linux a manual rm can fail silently if you don’t clear that first. Check with ls -l@ on macOS before deleting.

The reliable methods, in order of how persistent they are:

1. Mac/Linux quick path: Settings → System → toggle On-device AI off. PCWorld tested this and the weights.bin file vanished immediately on Mac.
2. Windows permanent block: Open regedit, go to HKEY_LOCAL_MACHINESOFTWAREPoliciesGoogleChrome (create the Google and Chrome keys if missing), add a DWORD named GenAILocalFoundationalModelSettings, set value to 1, restart. This approach was confirmed working as of May 2026.
3. Flags fallback: chrome://flags → search “optimization guide on device” → Disabled. Also disable “Prompt API for Gemini Nano” if it’s there.

Pro tip: The chrome://flags approach gets reset on major Chrome updates. If you’re on Windows and want a one-and-done fix, skip flags and go straight to the registry. On managed fleets, deploy via Group Policy instead – that’s the only method that survives every update channel cleanly.

Path B – Actually use the model

If you write code, you’ve got a free local LLM sitting on disk. The Prompt API lets you invoke Gemini Nano with no API key, no billing, no network round-trip. It’s been in Chrome since version 138 and is now in an Origin Trial for production sites.

Quick smoke test in DevTools:

const status = await LanguageModel.availability();
console.log(status); // "available" if ready

const session = await LanguageModel.create({
 systemPrompt: "You are a concise summarizer."
});
const result = await session.prompt("Summarize: ...");
console.log(result);

For streaming output, swap session.prompt() for session.promptStreaming(). The official docs note that the model is shared across origins – once installed, every AI-enabled page and extension on that machine benefits.

Common pitfalls (the ones tutorials skip)

The persistence layer trips people up. A few things to know before you waste an hour:

• Read-only deletion failures. If you nuke the folder and Chrome “keeps re-downloading,” check whether your delete actually succeeded – the read-only flag causes silent failures on Unix-like systems.
• The Mac toggle that isn’t there. 9to5Google testing found the On-device AI toggle missing from Settings → System on a MacBook running Chrome v147, while showing fine on Windows v147. If you can’t find it, you’re not crazy – the rollout is uneven. Use flags or wait for an update.
• The 10GB auto-purge. Per Chrome’s own developer docs, if free disk space drops below 10GB after install, the model gets removed automatically. It re-downloads when space returns. This is technically a soft uninstall lever – fill your drive, lose Nano – but obviously not a real solution.
• Hardware gate. Per official requirements as of mid-2026: more than 4GB VRAM, OR 16GB RAM + 4 CPU cores, plus 22GB free space, plus Windows 10/11, macOS 13+, Linux, or Chromebook Plus. No Android, no iOS. If your machine doesn’t meet those specs, the file won’t appear at all.

What to expect performance-wise

Nano is small. That’s the point and the limitation.

Treat it like the autocomplete tier of LLMs. You wouldn’t ask GPT-4 to do tab title suggestions and you wouldn’t ask Nano to write your tax return. Match the workload to the model.

When NOT to bother

Skip this entirely if:

• You build cross-browser web apps and need consistent behavior on Firefox/Safari – Nano is Chrome-only and the fallback dance is not free.
• Your laptop SSD is under 256GB. Four gigs sounds small until you also have Docker images, node_modules, and Xcode.
• You need accurate factual answers. Nano hallucinates more than its bigger siblings; metadata and precise knowledge are unreliable.
• You’re on a metered connection – the initial download is unmetered-only by design.

And honestly, if the consent angle bothers you ethically – that’s a fine reason on its own. Disable it. Move on.

FAQ

Is Gemini Nano actually private?

Inference runs locally, so yes – prompts you send through the Prompt API or built-in features like “Help me write” don’t leave your machine. The catch is the AI Mode button in the address bar isn’t part of that system; it sends queries to Google’s cloud. Don’t conflate the two.

I disabled it but the folder came back. Why?

Almost certainly one of three things happened: a Chrome update wiped your flag setting (they reset on major updates – this is the most common cause), you deleted the file but never disabled the underlying feature so Chrome just fetched it again, or a site called something like Summarizer.create() which re-triggers the download. The Windows registry fix or the Settings toggle – where it appears – survives all three scenarios. Use one of those, not flags.

Should I be worried about other browsers doing this?

Probably yes, eventually. Edge already ships with various AI hooks, and the on-device model trend isn’t going away. The Hanff investigation matters less because of Chrome specifically and more because it sets the precedent for whether “feature of an installed app” is a valid defense for silently writing multi-gigabyte binaries to user disks. That question hasn’t been answered in court yet.

Next step: open chrome://on-device-internals right now and look at what’s actually on your machine. If there’s nothing there, you’re not eligible – done. If there is, decide in the next 60 seconds: toggle off in Settings → System, or fire up DevTools and run LanguageModel.availability() to start building against it.