EdgeAI Kit 🧠📱

The missing UI toolkit for on-device LLMs in Flutter.

EdgeAI Kit fills the gap between flutter_gemma (inference engine) and what developers actually need: device capability checks, model management UI, onboarding flows, and ready-made chat/vision/tool-calling widgets.

Inspired by Google AI Edge Gallery (22.5k ⭐) — rebuilt for Flutter.

Screenshots

Features

Feature	Description
🔍 Device Capability Check	Detect RAM, GPU, storage, chipset. Classify device tier.
📋 Model Catalog	7 models with metadata, sizes, capability badges
🎯 Model Recommendations	Score models against device profile — "Great fit" / "Not enough RAM"
🚀 Onboarding Flow	First-launch: scan device → recommend → one-tap download → ready
💬 AI Chat	Multi-turn streaming conversation with markdown rendering
🖼️ Ask Image	Camera/gallery → model analyzes image on-device
🛠️ Tool Use	Real function calling — weather API, time, calculator with tool call visualization
🧠 Thinking Mode	Chain-of-thought reasoning visualization
⚙️ Prompt Lab	Test prompts with temperature, top-k, max tokens sliders
📊 Benchmark	Measure tokens/sec, TTFT on your device
💾 Storage Management	Delete models, storage usage bar, multi-model management
🎨 Material 3 Theme	Light + dark, ColorScheme.fromSeed, Google AI Edge Gallery style

Supported Models

Model	Size	Vision	Audio	Tools	Thinking
Gemma 4 E2B	2.4 GB	✅	✅	✅	✅
Gemma 4 E4B	4.3 GB	✅	✅	✅	✅
Gemma 3 1B	0.5 GB			✅
DeepSeek R1 1.5B	1.86 GB			✅	✅
Qwen3 0.6B	586 MB			✅	✅
Phi-4 Mini	3.94 GB			✅
SmolLM 135M	167 MB

Installation

Git dependency (add to your pubspec.yaml):

dependencies:
  edge_ai_kit:
    git:
      url: https://github.com/sumitvairagar/edge-ai-kit.git

Or clone and run the demo app:

git clone https://github.com/sumitvairagar/edge-ai-kit.git
cd edge-ai-kit
flutter run

Quick Start

import 'package:edge_ai_kit/edge_ai_kit.dart';

// Initialize
LLMService.initialize();

// Check device capabilities
final profile = await DeviceCapabilityChecker.check();
final recommendations = DeviceCapabilityChecker.recommendModels(profile);

// Download & load a model
await LLMService.install(ModelCatalog.all.first);
await LLMService.load(ModelCatalog.all.first);

// Use ChatView widget
ChatView(
  messages: messages,
  generating: isGenerating,
  onSend: (text) => handleSend(text),
  suggestions: ['Tell me a joke', 'What is Flutter?'],
)

Architecture

lib/
├── edge_ai_kit.dart              # Package barrel export
├── src/
│   ├── device/
│   │   └── device_checker.dart   # RAM/GPU/storage detection + recommendations
│   ├── models/
│   │   └── model_catalog.dart    # Model registry with metadata
│   ├── services/
│   │   ├── llm_service.dart      # flutter_gemma wrapper
│   │   └── tool_executor.dart    # Real tool implementations
│   ├── widgets/
│   │   └── chat_view.dart        # Shared chat UI component
│   └── theme/
│       └── app_theme.dart        # Material 3 theme
└── ui/screens/                   # Demo app screens (8 screens)

ChatView Message Types

ChatMessage.user('Hello!')                    // User bubble
ChatMessage.ai('Here is the **answer**...')   // AI bubble (markdown, streaming)
ChatMessage.thinking('Let me analyze...')     // Thinking block (collapsible)
ChatMessage.toolCall(name: 'get_weather',     // Tool call card (JSON)
  input: {...}, output: {...})
ChatMessage.loading()                         // Loading indicator

Device Capability Check

final profile = await DeviceCapabilityChecker.check();
print(profile.totalRamMB);      // 8192
print(profile.tier);             // DeviceTier.high
print(profile.gpuRenderer);      // "OpenGL ES 3.1+"

final recs = DeviceCapabilityChecker.recommendModels(profile);
for (final rec in recs) {
  print('${rec.model.name}: ${rec.fit} — ${rec.reason}');
  // Gemma 4 E2B: ModelFit.great — Great fit for your device
}

Benchmark Results (Gemma 4 E4B on Snapdragon)

Metric	Value
Tokens/sec	13.0 tok/s
Time to First Token	475 ms
Total Tokens	250
Total Time	18.6 sec

Requirements

• Flutter 3.41+
• Android 12+ / iOS 16+
• 4GB+ RAM recommended

Contributing

See CONTRIBUTING.md for guidelines on adding models, screens, and code style.

License

Apache 2.0 — see LICENSE.

Credits

Built with flutter_gemma by DenisovAV.