Repository files navigation

LLM Hub is an open-source Android app for on-device LLM chat and image generation. It's optimized for mobile usage (CPU/GPU/NPU acceleration) and supports multiple model formats so you can run powerful models locally and privately.

• 100% on-device processing - no internet required for inference
• Zero data collection - conversations never leave your device
• No accounts, no tracking - completely private
• Open-source - fully transparent

• GPU/NPU acceleration for fast performance
• Text-to-Speech with auto-readout
• RAG with global memory for enhanced responses
• Import custom models (.task, .litertlm, .mnn, .gguf)
• Direct downloads from HuggingFace
• 16 language interfaces

Quick Start

1. Download from Google Play or build from source
2. Open Settings → Download Models → Download or Import a model
3. Select a model and start chatting or generating images

Supported Model Families (summary)

• Gemma (LiteRT Task)
• Llama (Task + GGUF variants)
• Phi (LiteRT LM)
• LiquidAI LFM (LFM 2.5 1.2B + LFM VL 1.6B vision-enabled)
• Ministral / Mistral family (GGUF / ONNX)
• IBM Granite (GGUF)

Model Formats

• Task / LiteRT (.task): MediaPipe/LiteRT optimized models (GPU/NPU capable)
• LiteRT LM (.litertlm): LiteRT language models
• GGUF (.gguf): Quantized models — CPU inference powered by Nexa SDK; some vision-capable GGUF models require an additionalmmproj vision project file
• ONNX (.onnx): Cross-platform model runtime

GGUF Compatibility Notes

• Not all Android devices can load GGUF models in this app.
• GGUF loading/runtime depends on Nexa SDK native libraries and device/ABI support; on unsupported devices, GGUF model loading can fail even if the model file is valid.
• In this app, the GGUF NPU option is intentionally shown only for Snapdragon 8 Gen 4-class devices.

Importing models

• Settings → Download Models → Import Model → choose.task,.litertlm,.mnn,.gguf, or.onnx
• The full model list and download links live inapp/src/.../data/ModelData.kt (do not exhaustively list variants in the README)

Technology

• Kotlin + Jetpack Compose (Material 3)
• LLM Runtime: MediaPipe, LiteRT, Nexa SDK
• Image Gen: MNN / Qualcomm QNN
• Quantization: INT4/INT8

Acknowledgments

• Nexa SDK — GGUF model inference support (credit shown in-app About) ⚡
• Google, Meta, Microsoft, IBM, LiquidAI, Mistral, HuggingFace — model and tooling contributions

Development Setup

git clone https://github.com/timmyy123/LLM-Hub.gitcd LLM-Hub./gradlew assembleDebug./gradlew installDebug

To use private or gated models, add your HuggingFace token tolocal.properties (do NOT commit this file):

HF_TOKEN=hf_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

Save and sync Gradle in Android Studio; the app will readBuildConfig.HF_TOKEN at build time.

Contributing

• Fork → branch → PR. See CONTRIBUTING.md (or open an issue/discussion if unsure).

License

• MIT (see LICENSE)

Support

• Email:timmyboy0623@gmail.com
• Issues & Discussions: GitHub

Notes

• This README is intentionally concise — consultModelData.kt for exact model variants, sizes, and format details.

If you want, I can also add a short “Release notes / changelog” section and a quick performance guide for device profiles.