The Swiss Army Knife of On-Device AI — an offline-first mobile AI suite covering text generation, image generation, vision Q&A, and speech-to-text, with on-device RAG knowledge base and tool calling support.
Off Grid is a cross-platform (iOS / Android / Mac Catalyst) mobile offline AI application built with React Native. All inference runs locally on-device with zero data leaving the device.
Text Generation#
Supports any GGUF model (Qwen 3, Llama 3.2, Gemma 3, Phi-4, etc.) with streaming output, thinking mode, and Markdown rendering, achieving 15–30 tok/s on flagship devices, and supports user-supplied .gguf files.
Image Generation#
Integrates on-device Stable Diffusion with Snapdragon NPU acceleration on Android (5–10s/image) and Core ML acceleration on iOS, bundling 20+ models (Absolute Reality, DreamShaper, Anything V5, etc.) with AI Prompt enhancement.
Vision AI#
Supports real-time camera VQA with SmolVLM / Qwen3-VL / Gemma 3n for document analysis, scene description, and receipt recognition, with ~7s response on flagship devices.
Speech Input#
On-device Whisper for press-and-hold voice-to-text; audio never leaves the device.
Tool Calling#
Models supporting function calling can invoke built-in tools: web search, calculator, date/time, device info, knowledge base search, with auto tool loops and overflow protection.
On-Device RAG Knowledge Base#
Upload PDF/text documents, with MiniLM on-device Embedding, SQLite storage, and cosine similarity retrieval.
Document Analysis#
Attach PDFs, code files, CSVs in conversations with native PDF text extraction on both platforms.
Remote LLM Server Connection#
Connect to LAN OpenAI-compatible servers (Ollama, LM Studio, LocalAI) with auto model discovery, SSE streaming, Keychain-secured API key storage, and seamless local/remote model switching.
Architecture Highlights#
- Frontend: React Native (TypeScript 93.5%)
- Android Native Layer: Kotlin — llama.cpp / whisper.cpp / Stable Diffusion native module integration
- iOS Native Layer: Swift — CoreML Diffusion, PDFExtractor, DownloadManager, etc.
- RN Bridges: llama.rn, whisper.rn, op-sqlite
- Image Generation Backends: local-dream, ml-stable-diffusion, MNN
- Hardware Acceleration: Snapdragon NPU (Android), Core ML (iOS)
- CI/CD: GitHub Actions + SonarQube + Codecov
- E2E Testing: Maestro
Installation#
- Android: Google Play Store or GitHub Releases APK
- iOS: App Store (also runs on Apple Silicon Mac via Mac Catalyst)
- Build from Source: Node.js 20+, JDK 17 / Android SDK 36, Xcode 15+
MIT-licensed with 7 contributors, 79 releases, and 1,489+ commits.