react-native-foundation-models

React Native Nitro module for Apple's Foundation Models (Apple Intelligence). Enables AI capabilities in React Native applications using Apple's on-device model, with tool calling support and streaming responses on iOS 26.0+.

Requirements

• iOS 26.0 or later
• Apple Intelligence enabled in Settings > Apple Intelligence & Siri
• Compatible Apple devices (Apple Silicon Macs, recent iPhones/iPads)
• React Native with Nitro modules support

Installation

npm install react-native-foundation-models react-native-nitro-modules
yarn add react-native-foundation-models react-native-nitro-modules
bun add react-native-foundation-models react-native-nitro-modules

Usage

Basic Chat Session

import { LanguageModelSession } from 'react-native-foundation-models';

const session = new LanguageModelSession({
  instructions: 'You are a helpful assistant',
  useCase: 'general',
  guardrails: 'default',
});

// Stream responses
await session.streamResponse('Hello, how are you?', (response) => {
  console.log('Streaming response:', response);
});

Using React Hooks

import { useLanguageModel } from 'react-native-foundation-models';

function ChatComponent() {
  const { send, response, loading, error, isSessionReady } = useLanguageModel({
    instructions: 'You are a helpful coding assistant',
    onResponse: (response) => console.log('Got response:', response),
    onError: (error) => console.error('Error:', error)
  });

  const handleSendMessage = async () => {
    if (isSessionReady) {
      await send('Explain React hooks');
    }
  };

  return (
    <View>
      <Text>{response}</Text>
      <Button onPress={handleSendMessage} disabled={loading} title="Send" />
    </View>
  );
}

Tool calling

import { createTool, LanguageModelSession } from 'react-native-foundation-models';
import { z } from 'zod';

const weatherTool = createTool({
  name: 'weather_tool',
  description: 'Get current weather for a city',
  arguments: z.object({
    city: z.string(),
  }),
  handler: async (args) => {
    const response = await fetch(`/weather?city=${args.city}`);
    const res = await response.json();
    return res.data
  },
});

const session = new LanguageModelSession({
  instructions: 'You are a weather assistant',
  tools: [weatherTool]
});

API Reference

LanguageModelSession

Core class for managing AI conversations.

constructor(config?: {
  instructions?: string;
  tools?: Tool[];
  useCase?: 'general' | 'contentTagging';
  guardrails?: 'default' | 'permissiveContentTransformations';
})

The library now creates sessions with an explicit SystemLanguageModel, which lets you opt into Foundation Models use cases and guardrails from React Native.

useLanguageModel(config)

React hook for session management with automatic lifecycle handling.

Returns:

• session - The current session instance
• response - Latest AI response
• loading - Whether a request is in progress
• error - Any error that occurred
• send(prompt) - Send a message to the AI
• reset() - Reset the conversation state
• isSessionReady - Whether the session is ready to use

useStreamingResponse(session)

Lower-level hook for streaming responses.

checkFoundationModelsAvailability()

Check if Apple Intelligence is available on the device.

The returned object also includes:

• contextSize: current library context budget in tokens
• modelFamily: '26.0-26.3' or '26.4+' based on the OS version

getFoundationModelsModelFamily()

Returns the Foundation Models family for the current OS version: '26.0-26.3' or '26.4+'.

getFoundationModelsContextSize()

Returns the library's current context budget in tokens. Apple documents a 4,096-token context window for the on-device model in iOS 26.0 through 26.3, so the library currently surfaces that value until it can be wired to the native SystemLanguageModel.contextSize API with an iOS 26.4 SDK.

Foundation Models 26.4 Notes

• Apple’s current docs split the on-device model into two version families: 26.0-26.3 and 26.4+. The library exposes this as modelFamily so apps can version prompts when model behavior changes.
• Exact native token measurement with SystemLanguageModel.tokenCount(for:) is not compiled into this branch yet because the installed Xcode SDK here is iPhoneOS26.2, which does not expose that symbol. Once the project moves to an SDK that includes the 26.4 APIs, the JS surface can adopt the native token counter without another breaking API change.

Development

This project uses a workspace structure with:

• package/ - The nitro module source code
• example/ - Example app demonstrating usage

Setup

bun install
bun run build

Running the example

cd example
bun start    # Start Expo dev server
bun ios      # Run on iOS

Note: Android is not supported as this module requires Apple's Foundation Models framework.