jegly/Box

What it solves

Box provides a secure, fully offline environment for running advanced AI models directly on Android devices. It removes the need for cloud connectivity, ensuring privacy and enabling AI features like image generation, document analysis, and voice interaction even on de-Googled custom ROMs.

How it works

Box acts as a hardened, feature-rich interface for running models locally. It leverages hardware acceleration via LiteRT (formerly TensorFlow Lite) and llama.cpp to utilize the NPU, GPU, or CPU of modern mobile chipsets (such as Snapdragon, Google Tensor, and MediaTek). It supports GGUF model imports and includes a variety of specialized tools for vision, speech, and text processing, all while enforcing security measures like biometric locks and encrypted storage.

Who it’s for

This project is for Android users who want to run powerful, open-source AI models locally without relying on cloud services or Google's ecosystem. It is particularly useful for users of privacy-focused custom ROMs (like GrapheneOS or LineageOS) and those who need specialized accessibility tools like spoken camera assistance.

Highlights

• On-device Generative AI: Supports local image generation (FLUX.2/Bonsai), voice-to-speech chat, and music generation.
• Hardware Acceleration: Optimized for mobile NPUs and GPUs, including support for the latest Tensor and Snapdragon chips.
• Privacy-First: Operates entirely offline with no telemetry, featuring biometric app locks and encrypted chat histories.
• Accessibility Tools: Includes "Box Assist" for real-time, spoken camera feedback for blind and low-vision users.
• Extensive Model Support: Compatible with GGUF imports and a wide range of LLMs and vision models for document analysis and Q&A.