BreezeApp - AI on phone

English | 繁體中文

🔄 BreezeApp v2 Upgrade

We have upgraded to BreezeApp v2 with a new modular architecture using Git Submodules:

• BreezeApp-engine - Core AI engine (inference framework)
• BreezeApp-client - Engineering app (demo and testing platform)
• BreezeApp 🚧 (Coming Soon) - Production app (official consumer application)

✨ Key v2 Upgrade Features

1. Modernized Codebase (Java → Kotlin)

   • Migrated the entire codebase from Java to Kotlin for improved readability, safety, and modern Android development practices.

2. Modular Architecture: Decoupled AI Engine & UI

   • Fully separated the AI inference engine from the user interface, enabling independent development, easier maintenance, and better scalability.

3. EdgeAI SDK: Unified Developer-Friendly API

   • Introduced the EdgeAI SDK with standardized APIs for integrating AI features (chat, speech, vision) into any application.

Important

Migration Notice: The legacy breeze-app directory is deprecated. All new features and development are implemented in the v2 modular architecture.

Overview

BreezeAPP is a mobile AI application developed for both Android and iOS platforms. Users can download it directly from the App Store and enjoy a variety of features offline, including speech-to-text, text-to-speech, text-based chatbot interactions, and image question-answering. Currently, BreezeApp defaults to the Breeze 2 series of models developed by MediaTek Research and we provide simple installation instructions for changing your backend to Llama 3.2 See the Roadmap for additional planned integrations.

This project aims to create a community-driven platform for running AI capabilities locally on Android devices. Our goal is to provide a privacy-focused solution where all AI features work completely offline (airplane mode supported), ensuring your data never leaves your device.

Left: NPU Backend Right: CPU Backend

Speech-to-Text (In preparation)

Text-to-Speech

📧 Our Purpose of Open-Sourcing

We are MediaTek Research, a member of the AI Alliance.

Our main purpose for developing this application is to change the common perception that LLM capabilities can only be used on expensive devices and are limited by device providers. We aim to raise awareness and promote the concept that anyone can freely choose and run different LLMs on their phones.

Another main purpose of developing this application is to eliminate the obstacles developers may face when creating mobile AI applications by open-sourcing our Kotlin source code. We hope to inspire developers to create more innovative mobile applications.

If you are interested in BreezeApp, please contact us via [email protected]

🚀 Quick Start

📥 Download & Try

• Latest Release: Download APK
• First Launch: The app will automatically prompt you to download required AI models

👨‍💻 Developer Setup

Prerequisites: Git with submodule support

# Clone with submodules
git clone --recursive https://github.com/mtkresearch/BreezeApp.git

# Or initialize submodules after cloning
git submodule update --init --recursive

📖 Detailed Instructions: See our Setup Guide

Working with Submodules

This project uses Git Submodules for component management. See Submodules Guide for detailed instructions on:

• Initial setup and cloning
• Updating submodules
• Contributing to components
• Troubleshooting

✨ Features

🤖 AI Capabilities

• 💬 Chat: Text-based conversational AI
• 🗣️ Voice: Speech input/output support
• 👁️ Vision: Image understanding and Q&A
• 🔒 Privacy: 100% offline operation (airplane mode supported)

⚡ Backend Support

Feature	CPU Backend	MediaTek NPU	Status
LLM Chat	✅ Ready	✅ Ready	Stable
Vision (VLM)	🚧 Beta	❌ Planned	Testing
Speech Recognition	🚧 Beta	❌ Planned	Testing
Text-to-Speech	✅ Ready	❌ Planned	Stable

Legend: ✅ Production Ready | 🚧 In Development | ❌ Planned

🔍 Project Structure

BreezeApp/
├── BreezeApp-engine/          # 🔧 Core AI Engine (Git Submodule)
├── BreezeApp-client/          # 📱 Engineering/Demo App (Git Submodule)  
├── BreezeApp/                 # 🚧 Production App (Coming Soon)
├── docs/                      # 📚 Documentation
│   ├── setup/                 # Installation guides
│   ├── architecture/          # System design
│   ├── contributing/          # Contribution guidelines
│   └── i18n/                  # Internationalization
├── assets/                    # 🎨 Static resources
└── legacy/                    # ⚠️ Deprecated v1 code

🤝 Contributing

Contributions are welcome! See our Contributing Guide to get started.

📄 License

The license for this project is pending determination - see the LICENSE file for details.

🙏 Acknowledgments

• Executorch for LLM/VLM framework
• k2-fsa/sherpa-onnx for ASR/TTS capabilities
• MediaTek Research for core AI engines

🔗 Links

• Known Issues

📧 We are

MediaTek Research, a member of the AI Alliance.

Contact us: [email protected]

---

TWM Open Source Contributions Summary:

• fix: gradle failed due to ClassNotFoundException: org.gradle.wrapper.GradleWrapperMain
• fix: ClassCastException in MainActivity after closing the SettingsActivity
• Implement Instrumented Tests for LLMEngineService
• Implement Instrumented Tests for TTSEngineService, ASREngineService and VLMEngineService
• Code coverage report by unit test (Robolectric + Mockito) and instrumented test (JaCoCo)