Enabling Gemini Nano in Google Chrome

Google ships two different AI experiences inside Chrome. They share the "Gemini" name but work in completely different ways, and it is worth understanding the difference before you configure anything.

Gemini in Chrome is the cloud-based assistant you can open from the browser sidebar. It sends your page content and prompts to Google's servers and is managed through your Google account settings. Google Support has a full article about it here.

Gemini Nano is a lightweight language model that runs entirely on your own device. It powers the browser's built-in LanguageModel, Summarizer, Translator, Writer, Rewriter and Proofreader JavaScript APIs. No prompts, no page content and no personal data leave your computer. This is what the Admin Tools extension from Cloud Captains uses for its AI category.

This guide focuses on Gemini Nano. If you only wanted the cloud sidebar, you are already done once you sign into Chrome.

Why run AI on device?

Running AI locally has three practical advantages:

1. Privacy. Your text stays on the machine. For legal, confidential or internal content this removes a major compliance concern.
2. Cost. There is no API key to manage and no per-token billing. Once the model is downloaded it runs offline.
3. Latency. Responses arrive in milliseconds because there is no network round-trip.

The trade-off is that Gemini Nano is smaller than the frontier models that Google exposes through the API. It is excellent at summarising, rewriting, translation, structured text generation and simple chat. It is not designed for deep reasoning, long-context document analysis or code generation that competes with Gemini 3.0.

System requirements

Gemini Nano has non-trivial hardware requirements because the model is stored and executed on your machine.

Android, iOS and the Chrome for embedded Linux builds do not run Gemini Nano at the time of writing. Use a desktop build.

Step 1: verify your Chrome version

Open a new tab and enter:

chrome://settings/help

Chrome displays its current version and checks for updates. If you are below version 138, wait for the update to complete and restart the browser.

Step 2: enable the Chrome flags

Open a new tab and enter:

chrome://flags

Search for and change the following flags. Each flag has a dropdown on the right. Set every one of them to the value shown in the table.

The "BypassPerfRequirement" variant skips a performance gate that otherwise blocks some laptops with older GPUs. If your machine is not recent this is the setting that most often decides whether the model becomes available.

After changing each flag, a blue banner appears at the bottom of the page with the "Relaunch" button. Click it once you have changed all flags and Chrome restarts automatically. Your tabs are preserved.

Step 3: trigger the model download

The model is not downloaded until the browser knows you intend to use it. There are two ways to trigger the download.

Option A: use the Admin Tools extension

1. Open the extension popup and click AI Assistant in the sidebar.
2. Choose any AI tool, for example AI Chat or Summarizer.
3. Type a short test prompt ("Hello") and submit.
4. Chrome starts downloading the model in the background. The About page inside Admin Tools shows a live AI status panel with coloured dots. Yellow means "downloadable, in progress", green means "ready", red means "not available".

Option B: trigger from the DevTools console on any page

1. Open a plain page such as about:blank or any website.
2. Press F12 (Windows/Linux) or Cmd+Option+I (macOS) to open DevTools.
3. Select the Console tab.
4. Paste the following and press Enter:

await LanguageModel.availability()

5. If the response is "downloadable" or "after-download", paste:

const session = await LanguageModel.create();
await session.prompt("Hello in one word");

The first create() call kicks off the download. Progress can be monitored on the Chrome components page.

Step 4: monitor the download

Open a new tab and enter:

chrome://components

Look for the component called Optimization Guide On Device Model. If the status reads "Component not updated", click "Check for update". Once the download finishes the version field populates and the status switches to "up to date".

This component is roughly 2 GB and downloads in the background. Chrome throttles the download on metered connections, so switch to Wi-Fi if you are on a mobile hotspot.

Step 5: confirm everything works

Return to the Admin Tools popup, open About, and scroll to the AI status section. All seven APIs should show a green dot and the label "Available". If any dot is yellow or red, see the troubleshooting section below.

You can also verify from DevTools:

await LanguageModel.availability()

The result should be "available" (or "readily" on older Chrome builds).

Troubleshooting

"The flag is not present in chrome://flags"

You are on a Chrome version below 128 or on an unsupported platform. Update Chrome or switch to a desktop build.

"Availability returns 'unavailable' even after restart"

Three common causes:

• Insufficient disk space. Gemini Nano refuses to download when less than 22 GB is free. Clear some space and retry.
• GPU too old. The performance gate rejects very old integrated GPUs. Make sure you set "Enables optimization guide on device" to "Enabled BypassPerfRequirement", not just "Enabled".
• Managed Chrome profile. Enterprise or school accounts often have the flag policy-locked. Sign into Chrome with a personal account and retry.

"The model download never finishes"

Download progress is reported per Chrome component. If chrome://components shows 0.0.0.0 for "Optimization Guide On Device Model" after a long wait:

1. Click "Check for update" manually.
2. Close any tabs using the AI APIs.
3. Restart Chrome.
4. If still stuck, temporarily disable your VPN or ad-blocker. Some privacy extensions block the Google update servers.

"Responses are slow or Chrome freezes"

On a machine with only 8 GB of RAM the model competes for memory with normal browsing. Close unnecessary tabs and try again. Gemini Nano typical latency is 100 to 400 ms per streamed token. A full paragraph takes two to four seconds.

"My antivirus quarantined the model files"

This is rare but happens with some corporate endpoint security products. The model files live under the Chrome user data directory inside a folder named OptGuideOnDeviceModel. Add that path to your antivirus allowlist.

Using the APIs yourself

Once enabled you have access to the following browser globals in any page or extension:

• LanguageModel — conversational prompt API, streaming responses.
• Summarizer — paragraph summarisation with configurable type, format and length.
• Translator — language-to-language translation with auto-detection.
• Writer — long-form writing with tone, length and format options.
• Rewriter — rewrites input text with tone and length adjustments.
• Proofreader — grammar and spelling corrections with per-issue annotations.
• LanguageDetector — detects the language of a piece of text with confidence scores.

If you prefer to call them from your own code, Google publishes developer documentation on chromestatus and the built-in AI pages.

API maturity

Extensions bypass the Origin Trial registration requirement for the Prompt API. That is why Admin Tools can call LanguageModel directly without you adding a trial token.