An offline AI chat application for Android powered by MediaPipe LLM Inference. Chat with AI models entirely on-device - no internet required after downloading models.
| Model | Size | Capabilities |
|---|---|---|
| Smart Text Chat (Qwen 1.5B) | 1.57 GB | Conversations, coding, math, writing |
| High-Precision Text AI (Gemma3 1B) | 1.05 GB | Detailed text reasoning |
| Versatile Text AI (Gemma2 2B) | 2.71 GB | General text + Function Calling |
| Mobile Text Edge (Gemma2 2B) | 2.7 GB | Fast text on limited devices |
| Ultimate Vision Assistant (Gemma 3n) | 3.14 GB | Image analysis & visual Q&A |
| Thinking & Reasoning (DeepSeek R1) | 1.86 GB | Math, logic, chain-of-thought |
- Analyze images with the Gemma 3n vision model
- Ask questions about photos
- Image descriptions and visual understanding
With the "Versatile Text AI" model, you can:
- Open apps: "Open WhatsApp", "Launch YouTube"
- Browse web: "Go to google.com", "Search for weather"
- Make calls: "Call 123456789"
- Send SMS: "Send message to Mom"
- Set alarms: "Set alarm for 7am"
- Set timers: "Set timer for 5 minutes"
- Open settings: "Open WiFi settings"
- Take photos: "Take a photo"
- Play music: "Play some music"
- Conversation history with sessions
- Streaming responses (real-time typing effect)
- Dark theme UI
- Auto-scroll to new messages
- Session management (create, delete, switch)
- Customizable system prompt
- Clear chat history
- About & version info
Coming soon
- Android Studio Hedgehog (2023.1.1) or later
- Android device with Android 10+ (API 29+)
- At least 4GB free storage for models
- Clone the repository:
git clone https://github.com/yourusername/conversational-ai.git
cd conversational-ai-
Open in Android Studio
-
Build and run on your device:
./gradlew installDebug-
Grant storage permissions when prompted
-
Select a model and download it (requires internet for first download)
-
Start chatting offline!
Models are stored in:
/storage/emulated/0/Documents/NiqueWrld/models/
Models persist after app uninstall.
app/
├── src/main/java/com/niquewrld/conversationalai/
│ ├── MainActivity.kt # Main chat UI
│ ├── MainViewModel.kt # LLM inference & state management
│ ├── Model.kt # Model data class
│ ├── ChatSettingActivity.kt # Settings screen
│ ├── ChatHistoryActivity.kt # History screen
│ ├── Room/ # Database (sessions & messages)
│ ├── service/ # Model download service
│ ├── tools/ # Function calling executor
│ └── adapter/ # RecyclerView adapters
├── src/main/res/
│ ├── layout/ # XML layouts
│ ├── drawable/ # Icons & backgrounds
│ └── values/ # Colors, strings, themes
└── src/main/assets/
└── models.json # Available models config
- Language: Kotlin
- UI: Android Views + Material Design
- AI Inference: MediaPipe Tasks GenAI
- Database: Room
- Async: Kotlin Coroutines & Flow
- Architecture: MVVM with ViewModel
// MediaPipe LLM Inference
implementation("com.google.mediapipe:tasks-genai:0.10.29")
implementation("com.google.mediapipe:tasks-vision:0.10.29")
// Room Database
implementation("androidx.room:room-runtime:2.6.1")
implementation("androidx.room:room-ktx:2.6.1")
// Material Design
implementation("com.google.android.material:material:1.11.0")- First response is slower - model needs to warm up
- Smaller models = faster responses - try Gemma3 1B for speed
- Limit conversation length - start new sessions for best performance
- Close other apps - LLM inference uses significant RAM
This project is licensed under the MIT License - see the LICENSE file for details.
- MediaPipe for on-device ML inference
- Google Gemma for the language models
- Qwen for the Qwen 2.5 model
- DeepSeek for the R1 reasoning model
NiqueWrld
- Website: niquewrld.com
Made with ❤️ for offline AI enthusiasts
