jegly/Box

What is Box?

Box is an Android application that brings a large collection of on‑device AI capabilities to phones. It started as a fork of Google’s AI Edge Gallery and has diverged into a feature‑rich, security‑hardened suite that runs locally without needing an internet connection.

Core idea

On‑device inference – All models (large language models, vision models, diffusion generators, speech‑to‑text, text‑to‑speech, etc.) are stored and executed on the phone’s CPU, GPU, or dedicated NPU (Snapdragon, Google Tensor, MediaTek, etc.).
Privacy‑first – No cloud calls are required; chat history is encrypted, a biometric lock can protect the app, and Google‑owned telemetry (Firebase, Clearcut) has been removed.
One‑stop AI toolbox – The app bundles dozens of models for chat, image generation, image up‑scaling, document analysis, face recognition, music generation, multilingual speech, and a spoken‑camera assistant for low‑vision users.

Main features (as listed in the README)

Category	Highlights
LLM chat	Supports GGUF‑format models via `llama.cpp`; includes hybrid‑attention models (e.g., Qwen 3.5‑0.8B), reasoning‑focused models (Phi‑4‑mini‑reasoning), and large multimodal models (Gemma‑4‑12B). GPU acceleration via LiteRT and Vulkan is available where hardware permits.
Vision & multimodal	"Ask Image" with vision models such as SmolVLM2, InternVL3, and Gemini‑based models. Box Assist provides spoken‑camera assistance (object description, reading receipts, obstacle alerts) for blind/low‑vision users.
Image generation	On‑device diffusion models: Bonsai Image 4B (FLUX.2 klein), FLUX 2 klein, Z‑Image Turbo. Images are generated fully offline; optional up‑scaling with EDSR ×4.
Audio	Speech‑to‑speech chat (Voice Mode), multilingual STT via SenseVoice, on‑device TTS (Supertonic), and music/sound generation.
Document analysis	PDF/TXT ingestion, OCR, de‑blur, de‑noise, flattening of scanned pages, portrait‑sketch conversion.
Enhancements	Super‑resolution, face recognition, photo erase/in‑painting, biometric lock, encrypted SQLite (`SQLCipher`), tap‑jacking protection, offline‑only operation.
Hardware acceleration	LiteRT engine for NPU (Snapdragon 8 Gen 2/3/Elite, Google Tensor G5/G6, MediaTek), Vulkan GPU offload for GGUF models, and CPU fallback.
Internationalisation	UI available in English, French, Portuguese (BR), German, Simplified Chinese; multilingual speech models support 24 languages.
Accessibility	Colour‑blind‑safe themes, high‑contrast palettes, TalkBack compatibility, and the Box Assist spoken‑camera mode.

How to get it

Download the APK from the latest release page (e.g., Box v3.4.5).
For devices without Google services (LineageOS, GrapheneOS, CalyxOS) use the custom-rom-support APK.
Optionally install via Obtainium – add the repository URL (https://github.com/jegly/Box) and filter the APK name (Main or custom-rom-support).
The app includes an in‑app updater; you can also check for updates manually under Settings → Check for updates.

Licensing & provenance

License: Apache 2.0 (see LICENSE).
Origin: Fork of Google AI Edge Gallery (not affiliated with Google). The project has added 50+ features beyond upstream and stopped merging upstream changes.
Versioning: Upstream version shown as 1.0.15; Box’s own releases are tagged (e.g., v3.4.5).

Who might use it?

Power users who want a fully offline AI assistant on Android phones, especially those with high‑end NPUs.
Accessibility advocates looking for spoken‑camera assistance for blind or low‑vision users.
Developers / researchers interested in on‑device model deployment, hardware acceleration, or testing GGUF models without a cloud backend.
Privacy‑conscious users who want AI features without any telemetry.

Quick take‑away

Box turns a modern Android phone into a self‑contained AI workstation: chat, image creation, audio processing, and visual assistance all run locally, with optional hardware acceleration and a strong focus on security and offline operation. It is a genuine, actively maintained open‑source project in the mobile AI space.