RWKV App: A Flutter-based Local Large Language Model Chat Application

RWKV App is a cross-platform local LLM chat application built with Flutter, supporting Android, iOS, Windows, macOS, and Linux. It runs RWKV models offline, ensuring all data stays on the device to protect privacy. Key features include multi-model switching, local OpenAI-compatible API, text-to-speech, visual understanding, and more. It is open-source and caters to various scenarios like privacy-sensitive work and offline use.

The app emerges from the trend of deploying LLMs on edge devices. RWKV is an innovative model combining Transformer and RNN advantages: linear complexity (lower resource consumption), constant memory during inference, parallel training, and RNN-like token generation. These features make it ideal for edge devices, aligning with the app's 'privacy first' design.

1. Fully offline operation: All inference is local, no internet required.
2. Multi-model switching: Download from Hugging Face (0.1B-7B parameters; iPhone14+ runs 1.5B/2.9B smoothly).
3. Multi-round dialogue: Context-aware long conversations.
4. Local API server: OpenAI-compatible, integrating with other tools.
5. Text-to-speech (TTS) and visual understanding.
6. Dark mode support.

Built with Flutter for cross-platform compatibility. Download channels:

• Android: Google Play, GitHub, Hugging Face, 蒲公英
• iOS: App Store, TestFlight
• Windows: GitHub, Hugging Face, Microsoft Store
• macOS/Linux: GitHub, Hugging Face Specialized apps (Sudoku, Othello, Music) have limited platform support.

1. Flutter framework: Single codebase for all platforms, high performance.
2. Dart FFI & C++ engine: Efficient communication between Flutter and C++ for low-latency inference.
3. Multi-backend support: CPU (universal), GPU (acceleration), NPU/QNN (dedicated AI accelerators like Windows ARM64).

• Privacy-sensitive scenarios: Lawyers, doctors (data stays local).
• Offline environments: Planes, remote areas.
• AI democratization: Zero cost for ordinary users (no cloud subscriptions).
• Model testing: Researchers can switch models easily to compare performance.

Open-source (Apache 2.0 license). Community contributions welcome: code PRs, model adaptation, document translation (CN, TW, JP, KR, RU), issue feedback. Active QQ/Discord communities. Future plans: Integrate all functions into RWKV Chat; optimize performance; support larger models; expand multi-modal capabilities.

RWKV App represents a key direction in edge AI—bringing powerful LLMs to local devices while protecting privacy. It is an important part of the RWKV ecosystem and a reference for Flutter cross-platform development. With improved model efficiency and device computing power, local AI apps will have broader prospects.