jegly/Box

The most advanced, fully offline client-side AI suite on Android today.

What it solves

Box is a security-hardened, feature-rich Android application that enables the execution of various AI models entirely on-device, ensuring privacy and offline functionality. It serves as a comprehensive hub for local AI, providing tools for text generation, image generation, speech-to-text, text-to-speech, and computer vision, all without requiring a network connection.

How it works

Built as a fork of the Google AI Edge Gallery, Box utilizes a hybrid engine combining LiteRT and llama.cpp to run a wide array of models. It supports GGUF model imports and leverages hardware acceleration across a variety of mobile chipsets, including Google Tensor (TPU), Snapdragon, and MediaTek NPUs. The app provides a dedicated interface for managing these models and performing tasks like chat, image upscaling, and voice interaction.

Who it’s for

Android users who prioritize privacy and security (including those on custom ROMs like GrapheneOS or LineageOS) and AI enthusiasts who want to run powerful multimodal AI models locally on their mobile devices.

Highlights

  • Multimodal Capabilities: Supports text LLMs (Gemma, Qwen, Phi), image generation, image upscaling (super-resolution), and on-device music/sound generation.
  • Hardware Acceleration: Dedicated NPU support for Snapdragon, MediaTek, and Google Tensor (TPU) for faster local inference.
  • Privacy & Security: Features biometric lock, encrypted chat history via SQLCipher, and a "hard offline" mode.
  • Voice Mode: Integrated speech-to-speech AI chat using SenseVoice for STT and Supertonic for TTS.
  • Vision AI: On-device image recognition, face recognition (encrypted), and visual overlays for pose and face mesh detection.
  • Document Analysis: Ability to analyze PDFs and text files locally.
  • Extensibility: Support for llama.cpp and GGUF model imports, allowing users to bring their own models.