A mobile AI suite that never dials home

What it does

Off Grid is a React Native app that squeezes an unlikely amount of on-device AI into a single codebase. It runs GGUF text models such as Qwen, Llama, Gemma, and Phi; generates images with Stable Diffusion; answers questions about whatever your camera sees; and transcribes voice with Whisper—all without an internet connection. If you are on a local network, it can also hand off to OpenAI-compatible servers like Ollama or LM Studio, but its default posture is strictly offline.

The interesting bit

Instead of treating your phone like a thin client, the app acts like a self-contained workstation. It chunks and embeds your uploaded PDFs and text files using a bundled MiniLM model, stores vectors in SQLite, and retrieves them via cosine similarity—all locally. That is the kind of RAG pipeline usually bolted to a server, except here it works in airplane mode.

Key highlights

• Text generation hits 15–30 tok/s on flagship Snapdragon and Apple silicon, with image generation reaching 5–10s per image when NPU acceleration is available.
• Supports tool calling with built-in utilities including web search, calculator, and knowledge base search, plus runaway-prevention logic for automatic tool loops.
• Accepts custom .gguf files and can discover local-network LLM servers automatically, storing any API keys in the system keychain.
• Automatically rewrites simple user prompts into detailed Stable Diffusion prompts using the active text model before image generation.
• Vision inference via SmolVLM, Qwen3-VL, and Gemma 3n clocks in at roughly 7 seconds on flagship devices for document and scene analysis.

Verdict

Developers curious about edge AI or building privacy-first mobile experiences should study this to see how multiple native inference engines coexist inside a React Native shell. If you are after a bare-metal research framework or a server-side stack, this is just a very capable consumer app with its source open.