expo-ai-kit

On-device AI for Expo & React Native. Run LLMs locally — no API keys, no cloud, no cost.

Run Google's Gemma 4 (E2B / E4B), Apple Foundation Models, and ML Kit entirely on-device — chat, streaming, and downloadable models, all local.

• Private — inference never leaves the device
• Free — no API costs, rate limits, or subscriptions
• Native — Apple Foundation Models (iOS 26+), ML Kit (Android), Gemma 4 via LiteRT-LM
• Streaming — progressive token output with cancellation
• Model management — download, load, and switch Gemma 4 models at runtime

Install

npx expo install expo-ai-kit

Bare React Native projects: run npx pod-install afterwards. Android needs minSdkVersion 26 — set it via expo-build-properties in app.json.

Quick start

import { isAvailable, sendMessage } from 'expo-ai-kit';

if (await isAvailable()) {
  const { text } = await sendMessage([
    { role: 'user', content: 'What is the capital of France?' },
  ]);
  console.log(text); // "Paris"
}

Add a system prompt, or include system/assistant messages for multi-turn context. On-device models are stateless — pass the full message history on every call.

const { text } = await sendMessage(
  [{ role: 'user', content: 'Tell me a joke' }],
  { systemPrompt: 'You are a stand-up comedian.' }
);

Streaming

import { streamMessage } from 'expo-ai-kit';

const { promise, stop } = streamMessage(
  [{ role: 'user', content: 'Write a short story' }],
  (event) => setText(event.accumulatedText), // fired per token
);

await promise; // resolves with the final { text }
// stop();     // cancel at any point

Downloadable models (Gemma 4)

On Android, download and run Gemma 4 models on top of the built-in OS model.

import {
  getDownloadableModels, downloadModel, setModel, sendMessage,
} from 'expo-ai-kit';

// List models with their on-device status
const models = await getDownloadableModels();

// Download with progress, then activate
await downloadModel('gemma-e2b', {
  onProgress: (p) => console.log(`${Math.round(p * 100)}%`),
});
await setModel('gemma-e2b', { backend: 'auto' }); // 'auto' | 'gpu' | 'cpu'

// sendMessage / streamMessage now use the active model
const { text } = await sendMessage([{ role: 'user', content: 'Hi!' }]);

unloadModel() frees memory and reverts to the OS model; deleteModel(id) removes the file.

| Model | Params | Size | Platforms | |-------|--------|------|-----------| | gemma-e2b | 2.3B | ~2.6 GB | Android | | gemma-e4b | 4.5B | ~3.7 GB | Android |

iOS downloadable models are planned, pending LiteRT-LM Swift APIs from Google.

Platform support

| Platform | Engine | Models | |----------|--------|--------| | iOS 26+ | Apple Foundation Models | Built-in | | Android (supported devices) | ML Kit Prompt API | Built-in + downloadable Gemma 4 | | iOS < 26 / unsupported Android | — | isAvailable() returns false |

Requires Expo SDK 54+.

API

Inference

• isAvailable() → Promise<boolean>
• sendMessage(messages, options?) → Promise<{ text }>
• streamMessage(messages, onToken, options?) → { promise, stop }

Models

• getBuiltInModels() → Promise<BuiltInModel[]>
• getDownloadableModels() → Promise<DownloadableModel[]>
• downloadModel(id, { onProgress? }) / deleteModel(id)
• setModel(id, { backend? }) / unloadModel() / getActiveModel()

messages is { role: 'system' | 'user' | 'assistant'; content: string }[] and options accepts { systemPrompt?: string }. Model operations throw ModelError with a .code (e.g. MODEL_NOT_FOUND, DOWNLOAD_CORRUPT, INFERENCE_OOM). Full TypeScript definitions ship with the package.

Links

• Docs · npm · GitHub · Issues
• Apple Foundation Models · ML Kit GenAI · LiteRT-LM

License

MIT