If you run a recent version of Google Chrome on a desktop, there is a decent chance your browser has quietly downloaded a ~4GB artificial-intelligence model in the background. It is called Gemini Nano, and it is the engine behind Chrome’s new built-in AI features. The download is real — Snopes verified it — and it is worth understanding what it is, why it is mostly good, and where the legitimate concern lies.

What is actually on your machine

Gemini Nano is a compact, on-device language model that Chrome delivers through its component-updater system. The weights live in a file named weights.bin, inside a folder called OptGuideOnDeviceModel. You can check whether your browser has it — and its current size — by visiting chrome://on-device-internals in the address bar.

Per Chrome’s developer docs, the model powers a family of JavaScript APIs that web pages and extensions can call directly: a general LanguageModel (the “Prompt API”), plus Summarizer, Translator, Writer, Rewriter and Proofreader. It runs on Chrome for Windows 10/11, macOS 13+, Linux and Chromebook Plus — not yet on Android, iOS, or ordinary ChromeOS devices. The full APIs remain in an experimental/early stage, with broad stable availability targeted for Chrome 145–150 (late 2026 into 2027).

On-device means private — that part is genuinely good

The headline benefit is real: because the model runs locally, prompts and the text it processes do not have to be sent to a cloud server. For a browser that already sees a huge share of what people read and write, doing AI inference on the device — summarizing a page, translating text, proofreading a form — without shipping that content to Google’s servers is a meaningful privacy improvement over cloud AI. No round-trip, no server-side log of the prompt.

The fair concern: a 4GB install you didn’t really approve

The friction is consent and disk. Four gigabytes is not a rounding error. Consider the scale: Chrome holds roughly two-thirds of the global browser market (commonly cited around 66–68%, on the order of billions of users). If the model reaches even 500 million eligible desktops, that is about 2 exabytes of identical model weights sitting on consumer drives; reach a billion devices and it is ~4 exabytes. Most of those users never saw a clear “we’re about to download a 4GB AI model” prompt — it arrived as a background component update.

There is also a quieter shift worth naming: every browser becomes an AI runtime that any website can invoke. That is powerful for developers, but it also means a new local capability surface that security and privacy reviewers will need to reason about — rate-limiting, abuse of the on-device model by hostile pages, and fingerprinting based on model availability or version.

What to do about it

  • Check what you have: open chrome://on-device-internals to see if the model is present and how much space it uses.
  • Reclaim the space if you want: on metered or small-disk machines, you can manage Chrome’s optimization-guide / on-device model components; the model re-downloads only if a feature needs it.
  • Developers: treat the built-in APIs as progressive enhancement — feature-detect ('LanguageModel' in self), never assume availability, and don’t send anything to a page’s AI call you wouldn’t want processed locally.

Bottom line

Gemini Nano in Chrome is a real step toward private, local AI — and that is the right direction. The legitimate criticism is not the technology but the rollout: shipping a multi-gigabyte model to billions of machines deserves a clearer heads-up than a silent background update. Useful, mostly private, and a reminder that “your browser” now quietly includes an AI you didn’t explicitly install.

Tracking on-device AI and browser privacy — questions or corrections welcome via @mrtdnet on Telegram.