Gemini Nano for Android

A Flutter plugin to access Gemini Nano, Google's most efficient on-device AI model, directly through Android AI Core

This package enables offline-first, latency-free, and privacy-centric generative AI features by bridging Flutter with the native Android AICore (via ML Kit).

✨ Features

100% Offline Inference: Process data without an internet connection.
Privacy Focused: Data never leaves the device.
Zero Latency: No network round-trips; utilizes the on-device NPU (Pixel Tensor / Snapdragon).
Cost Efficient: Save on cloud API tokens by offloading simple tasks to the device.

📸 Example App

Example App

📱 Supported Devices

Gemini Nano via AI Core is currently available on select flagship Android devices, including but not limited to:

Google: Pixel 10, Pixel 10 Pro, Pixel 10 Pro XL, Pixel 10 Pro Fold, Pixel 9, Pixel 9 Pro, Pixel 9 Pro XL, Pixel 9 Pro Fold
Honor: Honor 400 Pro, Magic 6 Pro, Magic 6 RSR, Magic 7, Magic 7 Pro, Magic V3, Magic V5
iQOO: iQOO 13
Motorola: Razr 60 Ultra
OnePlus: OnePlus 13, OnePlus 13s, OnePlus Pad 3
OPPO: Find N5, Find X8, Find X8 Pro, Reno 14 Pro
POCO: POCO F7 Ultra, POCO X7 Pro
realme: realme GT 7 Pro, realme GT 7T
Samsung: Galaxy S25, Galaxy S25+, Galaxy S25 Ultra, Galaxy Z Fold7
vivo: vivo X200, vivo X200 Pro, vivo X Fold3 Pro, vivo X Fold5
Xiaomi: Xiaomi 15 Ultra, Xiaomi 15, Xiaomi 15T Pro, Xiaomi 15T, Xiaomi Pad Mini

Note: The user must have "AI Core" installed and updated via the Google Play Store.

🛠 Installation

Add the dependency to your pubspec.yaml:

dependencies:
  gemini_nano_android: ^0.0.1

⚙️ Android Configuration

Ensure your android/app/build.gradle defines a minimum SDK version compatible with modern AI features (API 26+ recommended):

android {
    defaultConfig {
        minSdkVersion 26
    }
}

🚀 Usage

1. Basic Text Generation

The simplest way to generate text.

import 'package:gemini_nano_android/gemini_nano_android.dart';

void generateText() async {
  final gemini = GeminiNanoAndroid();
  try {
    // Check if the device supports Gemini Nano
    bool isAvailable = await gemini.isAvailable();
    
    if (isAvailable) {
      List<String> results = await gemini.generate(
        prompt: "Explain quantum physics in 10 words.",
        temperature: 0.7, // Optional: Control creativity
        candidateCount: 1, // Optional: Number of responses
      );
      print("Gemini Nano says: ${results.first}");
    } else {
      print("Gemini Nano is not supported or not installed on this device.");
    }
  } catch (e) {
    print("Error generating text: $e");
  }
}

2. The "Hybrid Cloud-Fallback" Pattern (Recommended)

Since on-device models have limitations (context size, multimodal capabilities) or might not be downloaded yet, it is best practice to use a fallback strategy.

This example attempts to use Gemini Nano first (Free/Fast), and falls back to Firebase Vertex AI (Cloud) if it fails.

import 'package:gemini_nano_android/gemini_nano_android.dart';
// import 'package:firebase_ai/firebase_ai.dart'; // Uncomment if using Firebase

Future<String> smartProcess(String prompt) async {
  try {
    // 1. Try On-Device (Fast, Free, Private)
    print("Attempting local inference...");
    final gemini = GeminiNanoAndroid();
    final results = await gemini.generate(prompt: prompt);
    return results.first;
  } catch (e) {
    // 2. Fallback to Cloud (More powerful, costs money, requires internet)
    print("Local inference failed ($e). Switching to Cloud...");
    
    // final model = FirebaseVertexAI.instance.generativeModel(model: 'gemini-1.5-flash');
    // final response = await model.generateContent([Content.text(prompt)]);
    // return response.text ?? "Error in cloud generation";
    
    return "Fallback placeholder: Cloud generation would happen here.";
  }
}

3. Structured Output (JSON) for Receipt Processing

Gemini Nano is great at cleaning up OCR text.

Future<void> processReceiptText(String ocrRawText) async {
  final prompt = """
    Extract data from this receipt text. 
    Return ONLY a JSON with keys: 'total', 'date', 'merchant'.
    Text: $ocrRawText
  """;

  final gemini = GeminiNanoAndroid();
  final results = await gemini.generate(prompt: prompt);
  final jsonResult = results.first;
  // Parse jsonResult...
}

⚠️ Troubleshooting & Limitations

Model Not Downloaded: The first time an app requests the model, Android AI Core needs to download it (approx. 1GB+). This happens in the background. If you get a Model not found error, ensure the device is on Wi-Fi and charging, then try again later.

Multimodal Support: Currently, the Android AICore implementation for 3rd party apps is primarily Text-to-Text. To process images (like PDFs or receipts), use google_mlkit_text_recognition to extract text first, then pass that text to GeminiNanoAndroid.

Context Window: On-device models have smaller context windows than cloud models. Keep prompts concise.

📄 License

Distributed under the MIT License. See LICENSE for more information.