⚠️ BETA STATUS: This package is in beta. Core functionality including streaming is stable, but structured generation and tool calling features are still in development.

A Flutter package for integrating with Apple's Foundation Models framework on iOS and macOS devices. This package provides access to on-device AI capabilities through language model sessions, leveraging Apple Intelligence features.

• ✅ Cross-Platform Support: Works on both iOS 26.0+ and macOS 15.0+
• ✅ Persistent Sessions: Maintain conversation context across multiple interactions
• ✅ Streaming Responses: Real-time token streaming with delta updates
• ✅ Generation Options: Control temperature, token limits, and sampling strategies
• ✅ Transcript History: Access full conversation history with role-based entries
• ✅ Guardrail Levels: Configure content safety levels (strict/standard/permissive)
• ✅ Rich Responses: Get detailed responses with raw content and transcript metadata
• ✅ Security Features: Built-in prompt validation and injection protection
• ✅ Type-safe API: Built with Pigeon for reliable platform communication
• ✅ Privacy-First: All processing happens on-device with Apple Intelligence

• iOS: 26.0 or later
• macOS: 15.0 or later
• Flutter: 3.0.0 or later
• Dart: 3.8.1 or later
• Xcode: 16.0 or later
• Apple Intelligence: Must be enabled on device

Add this package to your pubspec.yaml:

dependencies:
  foundation_models_framework: ^0.2.0

Then run:

flutter pub get

To use this package on macOS, you must enable Apple Intelligence in your system settings. Go to System Settings → Apple Intelligence & Siri, enable Apple Intelligence, and allow the system to finish downloading the required on-device models. Without this step, the Foundation Models framework will not be available and the package will return an availability error.

In your ios/Podfile, ensure the iOS deployment target is set to 26.0 or higher:

platform :ios, '26.0'

In your ios/Runner.xcodeproj, set:

• iOS Deployment Target: 26.0
• Swift Language Version: 5.0

Physical Device (Recommended):

• Full Foundation Models functionality
• Real Apple Intelligence features
• Requires iOS 26.0+ device

Before using Foundation Models features, check if they're available on the device:

import 'package:foundation_models_framework/foundation_models_framework.dart';

final foundationModels = FoundationModelsFramework.instance;

try {
  final availability = await foundationModels.checkAvailability();
  
  if (availability.isAvailable) {
    print('Foundation Models is available on iOS ${availability.osVersion}');
    // Proceed with AI operations
  } else {
    print('Foundation Models not available: ${availability.errorMessage}');
  }
} catch (e) {
  print('Error checking availability: $e');
}

Create a session to interact with Apple's Foundation Models:

// Create a basic session
final session = foundationModels.createSession();

// Create a session with custom instructions and guardrails
final customSession = foundationModels.createSession(
  instructions: 'You are a helpful assistant. Keep responses concise.',
  guardrailLevel: GuardrailLevel.standard,
);

// Send a prompt and get a response
try {
  final response = await session.respond(prompt: 'Hello, how are you?');

  if (response.errorMessage == null) {
    print('Response: ${response.content}');

    // Access transcript history
    if (response.transcriptEntries != null) {
      for (final entry in response.transcriptEntries!) {
        print('${entry?.role}: ${entry?.content}');
      }
    }
  } else {
    print('Error: ${response.errorMessage}');
  }
} catch (e) {
  print('Failed to get response: $e');
}

// Don't forget to dispose when done
await session.dispose();

Control the model's generation behavior:

final session = foundationModels.createSession();

// Configure generation options
final options = GenerationOptionsRequest(
  temperature: 0.7,  // 0.0 = deterministic, 1.0 = creative
  maximumResponseTokens: 500,
  samplingTopK: 40,  // Top-K sampling
);

final response = await session.respond(
  prompt: 'Write a short story',
  options: options,
);

print('Generated: ${response.content}');

For single interactions, you can use the convenience method:

try {
  final response = await foundationModels.sendPrompt(
    'What is the weather like today?',
    instructions: 'Be brief and factual',
    guardrailLevel: GuardrailLevel.strict,
    options: GenerationOptionsRequest(temperature: 0.3),
  );

  if (response.errorMessage == null) {
    print('Response: ${response.content}');
  } else {
    print('Error: ${response.errorMessage}');
  }
} catch (e) {
  print('Failed to send prompt: $e');
}

For multi-turn conversations, reuse the same session:

final session = foundationModels.createSession();

// First interaction
var response = await session.respond(prompt: 'Tell me about Swift programming.');
print('AI: ${response.content}');

// Continue the conversation
response = await session.respond(prompt: 'Can you give me an example?');
print('AI: ${response.content}');

// Ask follow-up questions
response = await session.respond(prompt: 'How does that compare to Dart?');
print('AI: ${response.content}');

// Don't forget to dispose when done
await session.dispose();

For real-time token streaming:

final session = foundationModels.createSession(
  instructions: 'You are a helpful assistant',
  guardrailLevel: GuardrailLevel.standard,
);

// Stream tokens as they're generated
final stream = session.streamResponse(
  prompt: 'Write a detailed explanation of quantum computing',
  options: GenerationOptionsRequest(
    temperature: 0.7,
    maximumResponseTokens: 1000,
  ),
);

// Process tokens as they arrive
await for (final chunk in stream) {
  if (chunk.delta != null) {
    print('New tokens: ${chunk.delta}');
  }

  if (chunk.isFinal) {
    print('Complete response: ${chunk.cumulative}');
  }

  if (chunk.hasError) {
    print('Error: ${chunk.errorMessage}');
    break;
  }
}

You can cancel a stream at any time:

final stream = session.streamResponse(prompt: 'Long text generation...');

final subscription = stream.listen((chunk) {
  print('Received: ${chunk.delta}');

  // Cancel after receiving some content
  if (chunk.cumulative?.length ?? 0 > 100) {
    subscription.cancel(); // This will stop the stream
  }
});

The main class for accessing Foundation Models functionality.

• Returns: Future<AvailabilityResponse>
• Description: Checks if Foundation Models is available on the device
• Note: Returns true only if iOS 26.0+/macOS 15.0+ and Apple Intelligence is available

• Parameters:
  • instructions: Optional system instructions for the session
  • guardrailLevel: Content safety level (strict/standard/permissive)
• Returns: LanguageModelSession
• Description: Creates a new language model session with optional configuration

• Parameters:
  • prompt: The text prompt to send
  • instructions: Optional system instructions
  • guardrailLevel: Optional content safety level
  • options: Optional generation configuration
• Returns: Future<ChatResponse>
• Description: Convenience method to send a single prompt without managing a session

A persistent session for interacting with Apple's Foundation Models.

• Parameters:
  • prompt: The text prompt to send to the model
  • options: Optional generation configuration
• Returns: Future<ChatResponse>
• Description: Sends a prompt to the language model and returns the response

• Returns: Future<void>
• Description: Pre-warms the session to reduce first-token latency

• Returns: Future<void>
• Description: Disposes of the session and releases resources

• Parameters:
  • prompt: The text prompt to send to the model
  • options: Optional generation configuration
• Returns: Stream<StreamChunk>
• Description: Streams response tokens in real-time as they are generated

• bool isAvailable: Whether Foundation Models is available
• String osVersion: The OS version
• String? reasonCode: Structured reason code if unavailable
• String? errorMessage: Human-readable error message

• String content: The response content from the model
• String? rawContent: Raw response data
• List<TranscriptEntry?>? transcriptEntries: Conversation history
• String? errorMessage: Error message if the request failed

• String id: Unique identifier for the entry
• String role: Role (user/assistant/instructions/etc.)
• String content: The text content
• List<String>? segments: Individual text segments

• double? temperature: Controls randomness (0.0-1.0)
• int? maximumResponseTokens: Maximum tokens to generate
• int? samplingTopK: Top-K sampling parameter
• double? samplingProbabilityThreshold: Probability threshold for sampling

• strict: Maximum content safety
• standard: Balanced safety and flexibility
• permissive: More permissive content transformations

• String streamId: Unique identifier for the stream
• String? delta: New tokens in this chunk
• String? cumulative: All tokens received so far
• String? rawContent: Raw response data
• bool isFinal: Whether this is the last chunk
• String? errorCode: Error code if streaming failed
• String? errorMessage: Error message if streaming failed
• bool hasError: Convenience getter for error checking

The package handles errors gracefully and returns them in the response:

try {
  final response = await session.respond(prompt: 'Your prompt here');
  
  if (response.errorMessage != null) {
    // Handle specific errors
    switch (response.errorMessage) {
             case 'Foundation Models requires iOS 26.0 or later':
        print('Device not supported');
        break;
      case 'Foundation Models not available on this device':
        print('Apple Intelligence not available');
        break;
      default:
        print('Error: ${response.errorMessage}');
    }
  } else {
    print('Success: ${response.content}');
  }
} catch (e) {
  print('Unexpected error: $e');
}

• Foundation Models requires iOS 26.0 or later
• Only works on Apple Intelligence-enabled devices in supported regions

• All processing happens on-device using Apple's Foundation Models
• No data is sent to external servers
• Performance may vary based on device capabilities

• Always check availability before using features
• Handle errors gracefully for better user experience
• Consider providing fallback options for unsupported devices
• Test on actual devices with Apple Intelligence enabled

The package includes a complete example app demonstrating:

• Availability checking
• Session creation and management
• Prompt-response interactions
• Error handling

Run the example:

cd example
flutter run

Contributions are welcome! Submit pull requests for any improvements.

This project is licensed under the MIT License - see the LICENSE file for details.

See CHANGELOG.md for details about changes in each version.

For issues and questions:

• Create an issue on GitHub
• Check the example app for usage patterns
• Review Apple's Foundation Models documentation
• Check Apple's iOS 26.0+ release notes for hardware compatibility

This package implementation is based on Apple's Foundation Models framework:

• Apple Developer Documentation: Official API reference

Important: This package integrates with Apple's Foundation Models framework. Ensure you comply with Apple's terms of service and review their documentation for production use.