Box Brings Google's AI Edge Gallery Offline: Privacy-First Android Suite with Local Models
Key Takeaways
- ▸Box provides entirely offline, privacy-focused AI on Android, eliminating cloud dependency while maintaining fast inference
- ▸The project extends Google's AI Edge Gallery with encrypted chats, biometric security, and multi-engine support (LiteRT, llama.cpp, Stable Diffusion)
- ▸Supports full multimodal capabilities including voice-to-voice conversations, image generation, and vision AI with zero internet required
Summary
Box is a security-hardened, open-source fork of Google's AI Edge Gallery that brings comprehensive on-device AI capabilities to Android devices without requiring internet connectivity. Built by community developer thunderbong, Box enhances Google's original platform with encrypted conversations, biometric locks, and support for multiple inference engines including llama.cpp and Stable Diffusion. The app enables users to run multimodal AI tasks—including chat, voice conversations, image generation, document analysis, and vision AI—entirely offline, using either Google's LiteRT models (like Gemma 4) or custom GGUF models imported by users.
The project maintains compatibility with Google's original AI Edge Gallery while adding privacy-first features and expanded model support. Users can have flowing voice conversations with no wake words, accounts, or subscriptions required, with the AI responding sentence-by-sentence. A recent update introduced custom ROM support for alternative operating systems like LineageOS, GrapheneOS, and CalyxOS, though some features remain limited on these platforms due to missing system components like AICore and Text-to-Speech services.
- Available for both standard Android and custom ROMs, with separate builds for LineageOS, GrapheneOS, and CalyxOS
- Users can import any GGUF model or download optimized LiteRT models (including Gemma 4) directly in the app
