OGAM (Off Grid AI) is a mobile and desktop app that runs AI models on the device itself, on Android, iOS, and Apple Silicon Macs. Text chat, image generation, vision, speech-to-text, tool calling, and document analysis all run on local hardware without an account, and no data leaves the device.
Text generation runs GGUF models such as Qwen 3, Llama 3.2, Gemma 3, and Phi-4, or any GGUF file you supply, with streaming, a thinking mode, and markdown rendering. The app detects the device's hardware and picks the fastest backend: Adreno GPUs through OpenCL on Snapdragon, Metal on Apple Silicon, or the CPU. Image generation uses on-device Stable Diffusion (NPU-accelerated on Snapdragon, Core ML on iOS), vision uses models such as SmolVLM, Qwen3-VL, and Gemma 3n, and voice input uses on-device Whisper.
Models that support function calling can use built-in tools (web search, calculator, date and time, device info, and knowledge base search) in an automatic tool loop. Project knowledge bases chunk and embed uploaded PDFs and text on-device with a bundled MiniLM model and store them in SQLite. The app can also connect to OpenAI-compatible servers on the local network, such as Ollama, LM Studio, or LocalAI.
A paid Pro tier adds on-device text-to-speech, custom personas with persistent memory, actions drafted through Calendar, email, and MCP servers (such as Linear, Notion, and GitHub) that wait for your approval, and peer-to-peer sync between phone and Mac over your own network. The app is built on llama.cpp and whisper.cpp.
| Task | Flagship device | Mid-range device |
|---|---|---|
| Text generation (CPU) | 15-30 tok/s | 5-15 tok/s |
| Text generation (GPU, OpenCL) | 20-40 tok/s | n/a |
| Image generation (NPU) | 5-10 s | n/a |
| Image generation (CPU) | ~15 s | ~30 s |
| Vision inference | ~7 s | ~15 s |
| Voice transcription | Real time | Real time |
Features
- Local text generation: GGUF models such as Qwen 3, Llama 3.2, Gemma 3, and Phi-4, with streaming and a thinking mode
- Hardware acceleration: automatic choice of OpenCL GPU, Metal, NPU, or CPU backends, with manual override
- Image generation: on-device Stable Diffusion with real-time preview and 20+ models
- Vision: camera and image questions with SmolVLM, Qwen3-VL, and Gemma 3n
- Voice input: on-device Whisper speech-to-text with tap, auto, and hands-free modes
- Tool calling: web search, calculator, date and time, device info, and knowledge base search in an automatic tool loop
- Project knowledge base: PDFs and text embedded on-device and retrieved by cosine similarity from SQLite
- Remote servers: connects to OpenAI-compatible servers on the local network with model discovery and keychain-stored keys
- Prompt enhancement: the text model expands short prompts into detailed Stable Diffusion prompts
- Memory control: shows each resident model's RAM use, with per-model eject and loading policies
- Download manager: parallel, queued, resumable model downloads that survive backgrounding
- Pro tier: text-to-speech, personas, approval-gated actions through MCP servers, and device-to-device sync
