Skip to content

Chrome Local AI

Color mode

Chrome Local AI

Which built-in AI APIs your browser has, and the models on this device.

Demos of Chrome's built-in AI. A small language model, Gemini Nano, runs inside your browser: there is no server to pay for, nothing you write is uploaded, and once a model is downloaded it works without a connection.

Is your browser compatible?

Checked right now with the data your browser exposes. Nothing is sent anywhere.

Checking your browser…

A web page cannot read the free disk space (22 GB are needed) or the video memory (more than 4 GB). Chrome checks them itself when it downloads the model.

How to get started

  1. Open this site in Chrome on a computer (Windows, macOS or Linux). Other browsers do not have these APIs.
  2. In the table below, look at the state of each API. “Ready” means it can run now.
  3. If it says “Needs download”, press Prepare. Chrome downloads the model once; it can take a few minutes. Chrome only allows it after a click.
  4. Open any demo from the menu or the cards, type your own text or use the sample, and press the main button. The result appears as the model writes it.
  5. If something fails, read the frequently asked questions below, or open chrome://on-device-internals to see the state of the model.

Your browser

Seven APIs, each with its own job. Some demos use more than one.

APIStateOn this deviceAction
Prompt API LanguageModel
Ask the model anything: answers, writing, classification, data in a JSON shape, even images.
General-purpose generative AI
Stable
…
Summarizer API Summarizer
Shortens a text in a style and length you choose.
TL;DR, key points, teaser, headline
Stable
…
Translator API Translator
Translates between languages; each pair downloads its own small model.
On-device translation, one model per language pair
Stable
…
Language Detector API LanguageDetector
Says which language a text is written in, with a confidence.
Which language a text is in
Stable
…
Proofreader API Proofreader
Finds spelling, punctuation and grammar mistakes and explains them.
Origin trial; falls back to the Prompt API
Experimental
…
Writer API Writer
Writes a new text from an instruction, in a tone and length.
Deprecated; falls back to the Prompt API
Deprecated
…
Rewriter API Rewriter
Rewrites a text in another tone or length.
Deprecated; falls back to the Prompt API
Deprecated
…

“Prepare” downloads the model: Chrome only allows it after a click. For the details of the model, open chrome://on-device-internals and read the Broker State.

The language the on-device model is set up for (it takes English, Spanish and Japanese). It also decides the language of the example texts.

Requirements

AspectRequirement
Browser
Chrome on desktop (Windows, macOS, Linux). Not every API exists on mobile.
Model
Gemini Nano, downloaded once and run locally.
Network
Only for the first download of each model.
Languages
The main model takes English, Spanish and Japanese (more in newer Chrome versions); the Translator covers many more pairs.
Hardware
4 CPU cores and 16 GB of RAM, or a GPU with more than 4 GB of VRAM.
Storage
About 22 GB free for the models.

Demos

Each demo has a short guide at the top: what it does and how to use it.

  • Text toolkit

    One text, many actions: summarize, translate, proofread, simplify and rewrite, keeping every result to compare.

  • Prompt playground

    Send any instruction to Gemini Nano and watch the answer appear word by word.

  • Local chat

    A conversation with history, streaming, a stop button and several sessions you can fork.

  • JSON generator

    Describe the data you need and get it back as JSON that follows a schema.

  • Recipes

    Small ready-made prompts: classify, extract, tag, moderate, describe and ask questions of a text.

  • Synonym finder

    Synonyms of a word or phrase, with how each one differs.

  • Weather describer

    Turn raw weather data into a sentence a person would say.

  • Image analyzer

    Ask the model what is in a picture.

  • Summarizer

    Summarize a text as a TL;DR, key points, a teaser or a headline, in the length you choose.

  • Summary of summaries

    Split a long text into chunks, summarize each, then summarize the summaries.

  • Universal translator

    Paste text in any language: it is detected and translated to the one you pick.

  • Proofreader

    Spelling, punctuation and grammar fixes, each with its type and why.

  • Writer and rewriter

    Write from an instruction or rewrite a text in another tone and length.

  • Writing assistant

    An editor that corrects, changes the tone or simplifies the text you wrote.

Frequently asked questions

Setup, requirements and what to do when a model does not respond.

Getting started

Does anything I write leave my device?

No. The models run inside your browser, so the texts and images you use are not sent to any server. The only network use is the first download of each model.

Which browser and system do I need?

Chrome on desktop: Windows 10 or 11, macOS 13 or later, Linux, or ChromeOS on Chromebook Plus devices. Android and iOS are not supported by these APIs, and the Translator and the Language Detector are desktop only.

What hardware and disk space does it need?

These are the requirements Chrome lists. If the free space falls below 10 GB after the download, Chrome removes the model and downloads it again once the requirements are met.

  • At least 22 GB free on the disk that holds your Chrome profile.
  • A GPU with more than 4 GB of VRAM, or a CPU with 16 GB of RAM and 4 or more cores.
  • An unmetered connection for the first download.
How do I download a model?

Open the Status page and press Prepare on the API you want, or just use a demo. Chrome only starts the download after a click or a key press on the page, so it cannot happen by itself. Models are large and can take several minutes.

Does it work without internet?

Yes, once the model is downloaded. Each language pair of the Translator downloads its own small model the first time you use it.

When something does not work

A state says “Unavailable” or “Not in this browser”. What do I do?

“Not in this browser” means Chrome does not expose that API: use Chrome on desktop and keep it up to date. “Unavailable” means the API exists but the model cannot run on this device, usually because of the hardware or disk requirements.

These are the steps Chrome's guide recommends:

  • Restart Chrome.
  • Open chrome://on-device-internals and, in the Broker State tab, check that there are no errors.
  • Open the DevTools console and run LanguageModel.availability(). It should return available.
  • If it still fails, wait a little and repeat the steps.
The download does not start or does not finish.

Press Prepare, or the main button of a demo, on the page itself: the download needs a click. Then check that you have enough free disk space and an unmetered connection, and open chrome://on-device-internals to see the model and its size.

The answer repeats my prompt and ends with “End.Model:”.

That text is the request itself, sent back by the browser's model without writing any answer; it usually comes back in a few milliseconds. The prompt this site sends is correct, so the problem is in the model Chrome is running, probably a stand-in and not the real Gemini Nano. We have not been able to confirm the cause.

To check it:

  • Try the Prompt playground with a short text such as “Say hi”. If it also echoes the prompt, the model is the cause.
  • Open chrome://on-device-internals and check the model status: it should show a real model with a version and a size.
  • Restart Chrome and press Prepare again on the Status page.
  • Try the same page in Chrome stable on desktop.
The image analyzer returns an error or nonsense.

Image input is experimental in the Prompt API and depends on the Chrome version and on the model installed on your device. If the Prompt playground works with text, the model may simply not read images yet.

Why does a result say “fallback”?

The Proofreader is experimental, and the Writer and the Rewriter are deprecated. When Chrome does not have one of them, the site asks the Prompt API to do the job and marks the result as a fallback. The text is still valid, but it may differ from what the dedicated API would give.

About the demos

Which languages does the model understand?

The main model takes English, Spanish and Japanese; Chrome's guide adds German and French from Chrome 149. Choose the model language on the Status page: it also sets the language of the example texts. The Translator covers many more languages.

Why is the first answer slow?

The first request has to load the model into memory. The following ones start much sooner.

Do I need to change Chrome flags?

Chrome's getting-started guide lists no flags for these APIs, so try without them first. Experimental APIs, such as the Proofreader, can need extra steps: the setup guide has the current ones.

Not here? Read Chrome's setup guide

Resources