Brimkerndocs

Chrome extension · beta

Answers to your Google searches, computed on your device

The extension stops asking Google for its AI Overview and puts a button in its place. Click it, and a model running on your own GPU answers from the results already on the page, with its sources. No click, no computation.

Status: beta. Not on the Chrome Web Store yet: install it from this page.

A Google results page for “why is the sky blue”: above the results, the Brimkern panel shows a two-sentence answer, five numbered sources and the note “Computed on this device in 4.2 s”.
A real Google results page, 2026-11-19, Chrome on an M-series Mac: the answer, its sources, the time it took.

What it does

The AI Overview is not requested

Every web search goes out with Google’s own “Web” filter (udm=14), whose results page has no AI Overview. Hiding it would not be enough: by the time a page shows, Google has already generated it.

An answer only when you ask

Until you click, the extension makes no request and does not touch the GPU. Images, News and the other tabs work as usual.

Sources, always

Each answer lists the results it was built from, so you can check it in one click.

The same Google results page before the click: the Brimkern panel holds one line of explanation and a button, “Answer on this device”.
Before the click: the “Web” tab, no AI Overview, and a button. Nothing has been computed.

How it works

The model reads the titles and snippets of the results Google shows you, nothing else. It runs in a hidden page of the extension, so its weights are cached under the extension, not under google.com, and it closes itself after ten idle minutes to give the GPU memory back.

one search
google.com/search?q=…  + udm=14
   │ results page, no AI Overview
   ▼
[ Answer on this device ]  ← nothing happens until the click
   │ titles + snippets already on the page
   ▼
Brimkern  (hidden extension page)
   │
   ▼
your GPU  (WebGPU, hand-written WGSL kernels)
   └─▶ answer + sources, in the page

Fitted to your machine

The model is chosen once, at the first download, from what the browser says about the device and the connection. A model already on disk is always reused: no second download because the Wi-Fi got better.

WhenModelDownload
Data saver on, connection under 1.5 Mbit/s, or 4 GB of memory or lessQwen3 0.6B304 MB
OtherwiseQwen3.5 2B1.28 GB
No WebGPU, or a software-only adapterNo answer, and the panel says why: on the processor it would be very slow.

The first download is never silent: the panel shows the model, its exact size and where it comes from (Hugging Face), and waits for your click.

Measured

44 right answers in 48, none wrong

Qwen3.5 2B, on 24 real Google results pages (14 in French, 10 in English), each answered twice. The other 4 were partly right: vague, or a detail off.

First use: about 7 min 40 s

Click to answer, including the 1.28 GB download; it depends on your connection.

After a browser restart: 16 to 22 s

The first answer of a session moves the model from disk to the GPU.

Then: 5.3 to 5.9 s

Per answer, of which 4.2 to 4.6 s of computation.

Measured on 2026-11-15 on an M-series Mac in Chrome. Quality: pages captured from Google, answers read one by one (scripts/e2e/serp-quality.mjs). Times: the extension bench, on a fresh profile for the first use (scripts/e2e/extension-google.mjs).

Privacy and energy

The answer is built from what is already on your screen: the extension sends no extra request, except the one-time model download from Hugging Face. Google still receives your search; that is what a Google search is.

What it reads

On Google results pages only: your search and the titles, addresses and snippets of the results. Nothing on any other site.

Where it is processed

On your device, by the model running on your GPU. It never leaves the browser.

What it stores

The model weights, in the browser cache, so they are not downloaded twice. No search history, no answer, no account.

What it sends

Nothing to Brimkern: no analytics, no tracking. One request type only: the model download from Hugging Face, after you agree to it, which like any download reveals your IP address to Hugging Face.

Questions about this policy: github.com/RomainKH/Brimkern/issues

On energy, two facts and one commitment. The AI Overview is not requested, and nothing is computed unless you click. What a local answer costs in watt-hours has not been measured yet; we will publish that figure once it is measured, not an estimate.

Limits

A small model can be wrong

On our benchmark, Qwen3.5 2B gave no wrong answer, but some stay vague (the products to use where steps were asked). The light model got 2 answers in 48 wrong. The panel says so under every answer, next to the sources.

Chrome and Google only

Other Chromium browsers and other search engines are not tested yet.

The “Web” results page

Without the AI Overview also comes without the rich panels of the “All” tab (images, maps, knowledge panel). The other tabs stay one click away.

Install the beta

Until the Chrome Web Store release, the extension is installed by hand, in Chrome 116 or later. It takes a minute.

Download the beta (v0.1.0, .zip)
  1. Unzip the file, and put the folder somewhere it can stay: Chrome loads the extension from it at every start.
  2. Open chrome://extensions and turn on “Developer mode” (top right).
  3. Click “Load unpacked” and select the unzipped folder.
  4. Search on Google: the button appears above the results.

A beta installed this way does not update itself: for a new version, download it again, replace the folder and press the reload arrow in chrome://extensions. To remove it, click “Remove” there.

The zip is the output of the public build. To build it yourself:

git clone https://github.com/RomainKH/Brimkern && cd Brimkern
npm install
npm run build:extension
The engine and its models

Brimkern: open WebGPU engine, built by Romain Khanoyan. Local AI, WebGPU, on-device engines.