Language Model

Chrome AI provides on-device text generation via the built-in Prompt API (window.LanguageModel / Gemini Nano). Zero downloads, zero bundle weight, no API keys — everything runs locally in Chrome.

Browser support

The Prompt API is available on Chrome 138+ and Edge 138+ desktop. Firefox and Safari do not implement it; non-Chromium browsers report false from isPromptAPISupported() and the model rejects with a typed GenerationError.

Capability Detection

import { isPromptAPISupported } from '@localmode/chrome-ai';

if (isPromptAPISupported()) {
  // Safe to call chromeAI.languageModel()
}

isPromptAPISupported() checks the modern window.LanguageModel surface first, then falls back to the legacy self.ai.languageModel namespace used by Chrome 127–137 origin-trial builds. It never throws — it is safe to call in workers, Node, or non-browser environments.

Browser Support Matrix

Browser	Prompt API	Notes
Chrome 138+ desktop	Stable	Top-level `window.LanguageModel`
Edge 138+ desktop	Stable	Same Chromium surface
Chrome 127–137	Origin trial	Legacy `self.ai.languageModel` (auto-detected)
Chrome mobile (Android/iOS)	Not available	Use `@localmode/webllm` or `@localmode/wllama`
Firefox	Not available	—
Safari	Not available	—

Basic Usage

import { generateText } from '@localmode/core';
import { chromeAI } from '@localmode/chrome-ai';

const { text } = await generateText({
  model: chromeAI.languageModel({ systemPrompt: 'You are concise.' }),
  prompt: 'Explain quantum tunnelling in one sentence.',
});

console.log(text);

The factory accepts a ChromeAILanguageModelSettings object:

Setting	Type	Default	Description
`systemPrompt`	`string`	—	Prepended to every session as `initialPrompts[0]`
`temperature`	`number`	Chrome default	Sampling temperature (0–1)
`topK`	`number`	Chrome default	Top-K sampling cutoff
`contextLength`	`number`	`6144`	Soft documentation value for `model.contextLength`
`onProgress`	`(p: { loaded: number; total: number }) => void`	—	Forwarded to `monitor` for Gemini Nano download progress

For full API reference, options, and result types, see the Core Generate Text guide.

Streaming

import { streamText } from '@localmode/core';
import { chromeAI } from '@localmode/chrome-ai';

const { stream } = await streamText({
  model: chromeAI.languageModel(),
  prompt: 'Write a haiku about TypeScript.',
});

for await (const chunk of stream) {
  if (!chunk.done) process.stdout.write(chunk.text);
}

With the React hook:

import { useStreamText } from '@localmode/react';
import { chromeAI } from '@localmode/chrome-ai';

const model = chromeAI.languageModel();

export function Chat() {
  const { execute, data, isLoading } = useStreamText({ model });
  return (
    <div>
      <button onClick={() => execute('Tell me a joke')} disabled={isLoading}>Go</button>
      <pre>{data ?? ''}</pre>
    </div>
  );
}

Structured Output (`generateObject`)

Chrome's Prompt API does not yet expose constrained decoding — but generateObject() from @localmode/core works on every LanguageModel via prompt engineering + Zod validation + automatic retry.

import { generateObject, jsonSchema } from '@localmode/core';
import { chromeAI } from '@localmode/chrome-ai';
import { z } from 'zod';

const { object } = await generateObject({
  model: chromeAI.languageModel(),
  schema: jsonSchema(z.object({ name: z.string(), age: z.number() })),
  prompt: 'Extract: John is 30 years old',
});
// object: { name: 'John', age: 30 }

Warm-up

Eagerly create the underlying Chrome session so the next call has zero startup latency:

import { chromeAI } from '@localmode/chrome-ai';

const model = chromeAI.languageModel();
await model.warmUp();
console.log(model.isReady()); // true

With the React hook:

import { useModelWarmup } from '@localmode/react';
import { chromeAI } from '@localmode/chrome-ai';

const model = chromeAI.languageModel();

export function App() {
  const { isReady, warmUp } = useModelWarmup({ model });
  return <button onClick={warmUp}>{isReady ? 'Ready' : 'Warm up'}</button>;
}

Provider Fallback

Chain Chrome AI → WebLLM → wllama for graceful degradation:

import { generateText } from '@localmode/core';
import { chromeAI } from '@localmode/chrome-ai';
import { webllm } from '@localmode/webllm';
import { wllama } from '@localmode/wllama';

let model;
try {
  model = chromeAI.languageModel({ systemPrompt: 'You are helpful.' });
} catch {
  try {
    model = webllm.languageModel('Llama-3.2-1B-Instruct-q4f16_1-MLC');
  } catch {
    model = wllama.languageModel('Llama-3.2-1B-Instruct-Q4_K_M');
  }
}

const { text } = await generateText({ model, prompt: 'Hello' });

Error Reference

ChromeAILanguageModel throws GenerationError (from @localmode/core) with one of the following code values:

Code	When	Hint / Remediation
`chrome-ai-not-supported`	`window.LanguageModel` is missing entirely	Update to Chrome 138+ stable on desktop; verify the page is not in Incognito mode
`chrome-ai-model-not-available`	`availability()` returned `'unavailable'`	Enable `chrome://flags/#optimization-guide-on-device-model` and `chrome://flags/#prompt-api-for-gemini-nano`, then restart Chrome
`chrome-ai-download-required`	Gemini Nano is `'downloadable'` / `'downloading'` and `allowDownload` is not set	Pass `providerOptions: { chromeAI: { allowDownload: true } }` from a user-gesture handler
`chrome-ai-permissions-denied`	`LanguageModel.create()` threw `NotAllowedError`	Trigger generation from a user-gesture (click, tap, or keypress) on a same-origin page
`chrome-ai-multimodal-not-supported`	An `ImagePart` was supplied	Use `@localmode/webllm` (vision models) or `@localmode/wllama` (vision GGUFs); no resources are allocated before the rejection
`chrome-ai-quota-exceeded`	Input exceeded Gemini Nano's ~6K-token budget	Use `@localmode/webllm` with Llama-3.2 or larger for long-context generation
`chrome-ai-generation-failed`	Other Chrome-side failure	Inspect `error.cause` for the underlying error; consider falling back to another provider

Provider Options (`providerOptions.chromeAI`)

Key	Type	Default	Description
`topK`	`number`	—	Override per-call top-K (precedence over constructor)
`allowDownload`	`boolean`	`false`	Permit Chrome to download Gemini Nano on first use
`warnOnUnsupported`	`boolean`	`true`	Toggle the one-time `console.warn` for unsupported `topP`
`monitor`	`(m: EventTarget) => void`	—	Forwarded to `LanguageModel.create({ monitor })` for download progress

await generateText({
  model: chromeAI.languageModel(),
  prompt: 'Explain TLS',
  providerOptions: {
    chromeAI: {
      allowDownload: true,
      monitor: (target) => target.addEventListener('downloadprogress', (e) => console.log(e)),
    },
  },
});

Notes

topP is silently ignored. Chrome's Prompt API does not accept it. The first call with topP set logs a one-time console.warn; gate this with providerOptions.chromeAI.warnOnUnsupported.
stopSequences is post-processed. Chrome's API has no stop-token field, so the result text is truncated at the first stop sequence client-side.
maxTokens is informational. Chrome controls the real cap via the model's input quota (session.inputQuota).
Sessions are cached. ChromeAILanguageModel reuses one session per { systemPrompt, messages, temperature, topK } cache key. Changing any of those between calls destroys the previous session before creating a new one.

Language Model

On this page