Exploring Built-in AI Capabilities in Chromium-based browsers
Chromium's experimental built-in AI APIs run models directly in the browser, with no API keys or external services.
Chromium-based browsers now ship experimental built-in AI APIs. They run models directly in the browser, with no API keys and no external services.
Want a real-world application? My Browser Chat project is a chat app that runs entirely in the browser using these APIs.
Why would you want this?
- Your data never leaves the device
- No network latency
- No API costs
- Works offline once the model is downloaded
Want to try these demos? See how to set up Chrome’s built-in AI at the bottom of this article.
Prompt API
The Prompt API lets you generate text using a local language model. It’s like having a small ChatGPT running in your browser.
Try it out:
How to use it
// Create a sessionconst session = await LanguageModel.create({ monitor: (monitor) => { monitor.addEventListener("downloadprogress", (e) => { console.log(`Download: ${(e.loaded * 100).toFixed(1)}%`); }); },});
// Stream a responseconst stream = session.promptStreaming("Write a haiku about coding");const reader = stream.getReader();
while (true) { const { done, value } = await reader.read(); if (done) break; console.log(value);}
// Or get the full response at onceconst result = await session.prompt("Tell me a joke");console.log(result);Summarizer API
This one condenses long text into summaries. Pretty useful for articles or documents.
How to use it
const session = await Summarizer.create();
const text = `Your long text here...`;const summary = await session.summarize(text);console.log(summary);Rewriter API
The Rewriter API transforms text to match a tone: professional, casual, friendly, or confident.
How to use it
const session = await Rewriter.create();
const text = "I need you to review the document ASAP.";
// Make it more professionalconst professional = await session.rewrite(text, { tone: "professional" });console.log(professional);
// Or more casualconst casual = await session.rewrite(text, { tone: "casual" });console.log(casual);Language Detector API
Detects the language of any text and gives you a confidence score.
How to use it
const session = await LanguageDetector.create();
const result = await session.detect("Hola, ¿cómo estás?");console.log(result.detectedLanguage); // "es"console.log(result.confidence); // 0.95Translator API
Translates text between languages, all running locally.
How to use it
const session = await Translator.create({ sourceLanguage: "en", targetLanguage: "es",});
const translation = await session.translate("Hello, how are you?");console.log(translation); // "Hola, ¿cómo estás?"Using the Vercel AI SDK
If you want a more unified interface that works across different providers, you can use the Vercel AI SDK with the @browser-ai community providers.
This gives you three options:
@browser-ai/core: Chrome/Edge’s native built-in AI@browser-ai/transformers-js: Hugging Face models via Transformers.js@browser-ai/web-llm: open-source models via WebLLM
Installation
# For Chrome/Edge built-in AIpnpm add ai @browser-ai/core
# For Transformers.js (Hugging Face models)pnpm add ai @browser-ai/transformers-js
# For WebLLM (Llama, Qwen, Gemma, etc.)pnpm add ai @browser-ai/web-llm@browser-ai/core
This wraps Chrome’s Prompt API. Uses Gemini Nano on Chrome and Phi4-mini on Edge.
import { streamText } from "ai";import { browserAI } from "@browser-ai/core";
const result = streamText({ model: browserAI(), prompt: "Write a haiku about coding",});
for await (const chunk of result.textStream) { console.log(chunk);}@browser-ai/transformers-js
Runs Hugging Face models in the browser using WebAssembly. Works on all modern browsers.
import { streamText } from "ai";import { transformersJS } from "@browser-ai/transformers-js";
const result = streamText({ model: transformersJS("HuggingFaceTB/SmolLM2-360M-Instruct"), prompt: "Explain machine learning in one sentence.",});
for await (const chunk of result.textStream) { console.log(chunk);}@browser-ai/web-llm
Uses WebGPU for hardware-accelerated inference. Supports larger models like Llama 3 and Phi-3.5.
import { streamText } from "ai";import { webLLM } from "@browser-ai/web-llm";
const result = streamText({ model: webLLM("Qwen3-0.6B-q4f16_1-MLC"), prompt: "What are the benefits of WebAssembly?",});
for await (const chunk of result.textStream) { console.log(chunk);}Which one should you use?
| Provider | Best for | Browser support | Model size |
|---|---|---|---|
@browser-ai/core |
Fastest, Chrome/Edge users | Chrome 138+, Edge Dev | 2-4GB (built-in) |
@browser-ai/transformers-js |
Wide compatibility | All modern browsers | 270MB - 1GB |
@browser-ai/web-llm |
Better quality, larger models | Chrome/Edge with WebGPU | 500MB - 3GB+ |
Setting up built-in AI
If you want to try these APIs yourself, here’s how to set up Chrome.
Requirements
You’ll need:
- Chrome 138+ (desktop only)
- At least 22 GB of free storage
- Either a GPU with 4GB+ VRAM, or 16GB+ RAM with 4+ CPU cores
- Windows 10/11, macOS 13+, or Linux
Enabling the APIs
-
Update Chrome to version 138 or later
-
Go to
chrome://flagsand enable these flags:chrome://flags/#optimization-guide-on-device-modelchrome://flags/#prompt-api-for-gemini-nano-multimodal-input
-
Restart Chrome
-
Go to
chrome://componentsand click “Check for update” on “Optimization Guide On Device Model” -
Wait for the model to download (it’s about 2-4 GB)
Verifying the setup
Open your browser console and run:
const available = await LanguageModel.availability();console.log("Prompt API:", available);If it returns "available", you’re good to go. Other possible values are "downloadable", "downloading", or "unavailable".
For TypeScript projects, you can install the type definitions:
pnpm add -D @types/dom-chromium-aiThings to keep in mind
- The initial model download is 2-4GB
- Currently Chrome-only and experimental
- Needs decent hardware to run smoothly
Conclusion
Built-in AI is still experimental, but running models in the browser with no external dependencies is pretty cool. The privacy benefit alone makes it worth trying: your text never leaves the machine.
All the demos above are interactive. Try them and see how they perform on your hardware.
If you have questions or feedback, reach out.
Related Posts
The Agent Breakpoint
Phones made us build responsive apps. Agents are the next screen size, except there is no screen. What it takes to make a product agents can use: a CLI, an MCP server, and WebMCP.
Why System One models are a bigger deal than you think
A new kind of model answers typed questions with probabilities instead of writing text. I built a drawing game on one and tested 431 words against it. What it is, how it works, where it breaks.
AI Needs Tools, Not Just Intelligence
LLMs are prediction machines. Tools are what make them useful.