edge_gen_ai 0.3.0

A Flutter plugin for on-device generative AI, using Apple's Foundation Models (iOS) and Gemini Nano via ML Kit GenAI (Android).

edge_gen_ai

A Flutter plugin for on-device generative AI. It uses Apple's Foundation Models (iOS) and Google's Gemini Nano via ML Kit GenAI (Android) — no network calls, no cloud API keys, no data leaving the phone.

It uses the model the OS already ships with, so there's no model file to manage; downloadModel() just triggers the OS's own on-demand delivery of that shared system model when it isn't ready yet.

Important

Platform versions matter a lot here:

• iOS 26+ required, with Apple Intelligence enabled in Settings. iOS 27+ required for image input (Foundation Models Attachment, Beta).
• Android API 26+ on a device with Gemini Nano/AICore support (e.g. Pixel 8+, Samsung S23+). Android's backend is Beta — ML Kit GenAI APIs have no SLA or backward-compatibility guarantee. See ML Kit GenAI.

Use checkAvailability() to detect unsupported OS versions, disabled Apple Intelligence, or missing AICore support at runtime.

Features

Class	Task	Details	Screenshot
EdgeGenAIPrompt	generateContent()	Free-form prompt, streamed response, optional image input and conversation memory.
EdgeGenAISummarizer	summarize()	Summarizes text as bullet points.
EdgeGenAIProofreader	proofread()	Fixes grammar, spelling, and punctuation.
EdgeGenAIRewriter	rewrite()	Rewrites text in a chosen EdgeGenAIRewriteStyle.
EdgeGenAIImageDescriber	describeImage()	Describes an image.
EdgeGenAITool	Tool (function) calling	Runs a Dart callback when the model requests one of your tools.

Tool (function) calling

Add EdgeGenAITools to an EdgeGenAIPrompt. When the model uses a tool, the plugin runs your Dart callback and gives its result back to the model.

Caution

Function calling differs by platform. iOS uses the Foundation Models framework's native tool-calling support. The Android model does not currently support function calling through ML Kit GenAI, so this plugin adds it manually: it asks the model to return a JSON tool call, parses that response, runs the Dart callback, and sends the result back to the model. This Android path is best-effort—the model can return malformed JSON or answer without calling the tool. Test every function-calling flow carefully on your supported Android devices and model versions.

import 'package:edge_gen_ai/edge_gen_ai.dart';

Future<void> runFunctionCallingExample() async {
  final calculator = EdgeGenAITool(
    name: 'multiply_numbers',
    description: 'Multiplies two numbers and returns the result.',
    parameters: [
      EdgeGenAIToolParameter(
        name: 'left',
        description: 'The first number to multiply.',
        type: EdgeGenAIToolParameterType.number,
      ),
      EdgeGenAIToolParameter(
        name: 'right',
        description: 'The second number to multiply.',
        type: EdgeGenAIToolParameterType.number,
      ),
    ],
    onCall: (arguments) async {
      final left = arguments['left'];
      final right = arguments['right'];

      if (left is! num || right is! num) {
        return 'Both left and right must be numbers.';
      }

      // This is ordinary Dart code. It can also call your database,
      // device APIs, or a remote service and return their result.
      return '${left * right}';
    },
  );

  final assistant = EdgeGenAIPrompt(tools: [calculator]);

  await for (final response in assistant.generateContent(
    'Use the calculator to multiply 17 by 24.',
  )) {
    // Each event contains the complete response generated so far.
    print(response);
  }
}

You can register multiple tools on the same EdgeGenAIPrompt, including tools with no parameters or optional parameters. For a complete Flutter UI with several examples, see function_calling_page.dart.

Every class exposes checkAvailability() and downloadModel() alongside its task method. On Android these map to ML Kit GenAI's dedicated APIs; on iOS they're task-specific prompts to the same Foundation Model that backs EdgeGenAIPrompt.

• Availability + download: checkAvailability() reports ready / downloadable / not-enabled / unsupported. downloadModel() streams download progress on Android; on iOS it completes immediately (nothing to download).
• Conversation memory: EdgeGenAIPrompt(useMemory: true) remembers prior turns; resetConversation() starts over. Stateless by default.

Usage

import 'package:edge_gen_ai/edge_gen_ai.dart';

final prompt = EdgeGenAIPrompt();

// 1. Check whether the on-device feature is ready.
final availability = await prompt.checkAvailability();

// 2. If needed, download it (no-op on iOS).
if (availability == EdgeGenAIAvailability.downloadable) {
  await for (final progress in prompt.downloadModel()) {
    print('${progress.status}: ${progress.bytesDownloaded ?? ''}');
  }
}

// 3. Generate content. Each event is the full text generated so far.
await for (final chunk in prompt.generateContent(
  'Write a 3 sentence story about a magical dog.',
  options: EdgeGenAIGenerationOptions(temperature: 0.8, maxOutputTokens: 256),
)) {
  print(chunk);
}

// Optionally attach a single image (encoded bytes, e.g. PNG or JPEG).
await for (final chunk in prompt.generateContent(
  'What is in this picture?',
  image: imageBytes,
)) {
  print(chunk);
}

Hold a memory-enabled conversation across calls:

final chat = EdgeGenAIPrompt(useMemory: true);
await for (final chunk in chat.generateContent('My name is Alex.')) {}
await for (final chunk in chat.generateContent('What is my name?')) {
  print(chunk); // remembers "Alex"
}
await chat.resetConversation(); // start fresh

Use the task-specific features (each has its own checkAvailability() / downloadModel(), exactly like EdgeGenAIPrompt):

final summary = await EdgeGenAISummarizer().summarize(longArticle);

final corrected = await EdgeGenAIProofreader().proofread('the quick brown fox jumsp');

final formal = await EdgeGenAIRewriter().rewrite(
  'hey, meeting is off',
  style: EdgeGenAIRewriteStyle.professional,
);

final description = await EdgeGenAIImageDescriber().describeImage(imageBytes);

See the example app for a chat UI and a text-tools demo built on top of this API.