Skip to content

Chrome's Built-in AI (Gemini Nano): What It Can and Can't Do

Chrome now ships a small AI model that runs on your own computer. What it's good at, where it falls short, which APIs are ready, and when to use your own key.

Computing··10 min read

Desktop Chrome can now run a language model on your own computer. Websites reach it through a handful of JavaScript APIs, and the model behind most of them is Gemini Nano, the smallest member of Google's Gemini family. Chrome downloads it once, and after that, in Google's words, no data is sent to Google or any third party when you use it.

That means summaries, rewrites and proofreading with no account, no API bill and no text leaving the machine. It's also easy to expect too much of it. Here's what it does well, where it struggles, and when a bigger model with your own key makes more sense.

What "built-in AI" actually means

Chrome exposes a set of task APIs, each built for one job, plus one general-purpose API (the Prompt API) that takes free-form instructions. According to Chrome's get-started guide, the Translator and Language Detector use small expert models trained for that one task, and everything else runs on the Gemini Nano language model.

Here's where each API stands for ordinary web pages, from Chrome's API status table as of October 2026:

API What it does Model Status on the web
Summarizer Key points, TL;DR, teaser or headline for a text Gemini Nano Stable since Chrome 138
Translator Translates between language pairs Expert model Stable since Chrome 138
Language Detector Guesses the language of a text Expert model Stable since Chrome 138
Prompt API Free-form instructions, plus image and audio input Gemini Nano Stable since Chrome 148 (extensions had it from 138)
Writer Drafts new text from a short brief Gemini Nano Developer trial
Rewriter Makes text shorter, longer, more formal or more casual Gemini Nano Developer trial
Proofreader Fixes grammar, spelling and punctuation Gemini Nano Developer trial

"Developer trial" means a site can't count on these being switched on for its visitors, so a well-built page falls back to the Prompt API, which can do the same jobs from a plain instruction. Google says it is working to standardise the APIs across browsers, but for now they're Chrome's.

What it's good at

Google's examples for these APIs are telling: summarise a meeting transcript, pull key points out of a long article, make an email more polite, translate a post on request, correct a comment before it's sent. They're all transformations of text you already have. That's where a small local model earns its keep.

  • Summarising your own notes. The Summarizer offers four shapes: key points (3, 5 or 7 bullets), a TL;DR (1, 3 or 5 sentences), a teaser, or a one-line headline.
  • Rewriting and proofreading. The model only has to rearrange or correct what's there, not know anything new.
  • Translating. Chrome's Translator documentation lists 39 languages, with language packs downloaded on demand.
  • Detecting a language before translating. Chrome warns it's unreliable on single words and very short phrases.
  • Turning a short description into structured text, such as a small diagram or a checklist, when the output can be checked by a parser afterwards.

The common thread: the input contains everything the answer needs. A web.dev guide by Maud Nalpas adds the other draws of client-side AI: low latency, lower server costs, no API keys, more privacy and offline access.

What it can't do (yet)

Run on most phones, or on modest laptops

The models that power these APIs aren't available on mobile. Chrome for Android and iOS are not supported, and neither are ChromeOS devices other than Chromebook Plus. On desktop, Chrome's hardware requirements are:

Requirement What Chrome asks for
Operating system Windows 10 or 11, macOS 13 (Ventura) or later, Linux, or ChromeOS 16389.0.0+ on Chromebook Plus
Free storage At least 22 GB on the drive that holds your Chrome profile
Graphics or processor A GPU with more than 4 GB of video memory, or 16 GB of RAM and at least 4 CPU cores
Network An unmetered connection for the first download

The 22 GB is headroom: Chrome says the model itself is "significantly smaller", and chrome://on-device-internals shows its current size and status. If free space drops below 10 GB, Chrome removes the model and fetches it again later.

The Translator and Language Detector aren't tied to that list; Chrome says only that they work in desktop Chrome and not on mobile.

Know much about the world

Small is the point, and small has costs. Google's 2023 Gemini technical report gave the original Nano models 1.8 billion and 3.25 billion parameters, sized for devices. Chrome's docs don't give a parameter count for today's version, but it's a laptop-sized model, not a data-centre one.

So don't ask it for facts, citations, maths or code you can't check. Give it the material and ask it to work on that. A summary of your meeting notes is a good request; "what are the side effects of this medicine?" is not.

Handle long documents in one go

Each Prompt API session reports its context window and how much is used; Chrome doesn't promise a fixed size. When a request won't fit, the call fails with a QuotaExceededError that says how many tokens were asked for and how many were available. Chrome's own guidance suggests a "summary of summaries" for long content: summarise sections, then summarise the summaries.

Speak every language

Chrome's docs say that from Chrome 149 the Gemini Nano APIs (Summarizer, Writer, Rewriter, Proofreader and the Prompt API) support English, Spanish, Japanese, German and French for input and output. Hindi notes or a Polish email may be refused, even though the separate Translator can handle those languages.

Start without a click

The first time a page needs the model, Chrome has to download it, and it will only start that download in response to something you did, such as pressing a button. A site can't quietly start the download while you read.

How we use it on Binary Decimal

Two of our tools use the built-in model by default. Neither needs an account, and neither sends your text anywhere unless you choose your own key.

The Markdown editor has five AI actions on a selection or the whole page:

  • Summarise: key points or a TL;DR, which you can place at the top as a summary block, insert below, or copy.
  • Rewrite: shorter, longer, friendlier or more formal.
  • Proofread: every suggested change is listed separately, so you can accept the fixes you want and skip the rest.
  • Translate: through Chrome's Translator, with the source language detected for you.
  • Write: a first draft from a short brief.

The editor uses the dedicated API when Chrome offers it and the Prompt API when it doesn't. Summaries go to the Prompt API first, because in our testing it gave tighter summaries of notes than the Summarizer did. Long text is split by paragraph and sentence so each part fits the model; past six parts, the editor asks you to select a smaller piece rather than quietly truncating.

The Mermaid diagram editor uses the Prompt API for three jobs: Describe (type "a sign-in flow with a retry" and get a flowchart), Explain (a plain-language reading of the diagram that's open) and Fix (repair a diagram that won't parse). Every answer goes through Mermaid's own parser before you see it; if it doesn't parse, the model gets one retry with the error, and if that fails you see the error, not a broken diagram. A small model works well when something else can check its output.

The first time you use either, the panel asks before downloading the model and then shows a progress bar. After that there's no wait for a download, and it works without a connection.

When to bring your own key

Settings → AI lets you swap the built-in model for a bigger one from OpenRouter, OpenAI, Anthropic or Google Gemini, with your own API key (or, for OpenRouter, by signing in). Every AI feature on the site then uses that model instead.

Reach for your own key when:

  • You're on a phone, or a computer that doesn't meet the requirements. The built-in option simply won't be there.
  • The text isn't in one of the five supported languages and you want to summarise or rewrite it, not just translate it.
  • The job needs reasoning or knowledge, such as restructuring a long document, drafting from a vague brief, or turning a complicated process into a diagram.
  • The document is long and you'd rather not work through it a section at a time.

Stay with the built-in model when the text is private and the job is a straightforward transformation: tidy up an email, summarise your own notes, fix the grammar in a paragraph.

How your key is handled, and what changes:

  • The key stays in your browser: for this tab only by default, or on this device if you tick "Remember on this device". It's never synced, never put in a backup or a link, and never sent to Binary Decimal.
  • When you use an AI feature, the text you ask it to work on goes from your browser straight to the provider you picked, under their terms. We don't proxy it, and the AI panel says where it's going. That's the trade: a bigger model, but your text leaves your device.
  • While a key is stored, the site doesn't load Google Analytics, since any script on the page could read browser storage.
  • You pay the provider for what you use. Requests are billed in tokens; What LLM tokens really cost explains how to estimate a bill before it arrives.

Check whether your computer can use it

Select a paragraph in the Markdown editor in desktop Chrome and choose an AI action: if the model can run, you'll be offered the download. Settings → AI says "not available in this browser" when it can't, and chrome://on-device-internals shows the model's status.

The short version

  • Chrome's built-in AI runs Gemini Nano on your computer. After a one-time download, your text stays on the device and works offline.
  • Summarizer, Translator and Language Detector are stable since Chrome 138, and the Prompt API since Chrome 148. Writer, Rewriter and Proofreader are still in developer trials.
  • It needs desktop Chrome, 22 GB free, and either a GPU with more than 4 GB of video memory or 16 GB of RAM and 4 cores. No phones yet.
  • It's good at rewriting, summarising, proofreading and translating text you give it. It's poor at facts and long documents, and the language model works in five languages.
  • Bring your own key when you need a bigger model, other languages or a phone. Your text then goes to the provider you chose.

Sources

aichromegemini nanoon-device aiprivacyllm

Keep reading

Computing · Jul 18, 2026 · 2 min

What LLM Tokens Really Cost (and How to Estimate Your Bill)

Tokens, context windows, input vs output pricing: how large language model costs add up, and how to estimate your monthly bill before it surprises you.

Computing · Oct 7, 2026 · 9 min

Common JSON Errors and How to Fix Each One

Trailing commas, single quotes, comments, NaN, escaped strings and big numbers that change: why JSON rejects each one, and the fix that works.

Computing · Oct 7, 2026 · 11 min

Mermaid Flowcharts and Sequence Diagrams: A Practical Guide

Write Mermaid flowcharts and sequence diagrams that render first time: shapes, arrows, subgraphs, alt and loop blocks, and the errors that break them.