Off Grid AI: On-Device AI Suite for Phone and Mac
off-grid-ai/off-grid-ai-mobile
Off Grid AI is an MIT-licensed React Native app that runs text chat, image generation, vision AI, and speech-to-text fully on-device across Android, iOS, and Apple Silicon Macs. Reach for it if you want one app that replaces several separate local-AI tools and can tolerate an experimental NPU path; skip it if you need voice talkback without paying for the $69 Pro tier.
Understanding Off Grid AI
Off Grid AI is a React Native application for Android, iOS, and Apple Silicon Macs that bundles GGUF-based text chat, on-device Stable Diffusion image generation, camera-based vision AI, and Whisper speech-to-text into one offline suite. Every model runs locally on the phone's or Mac's CPU, GPU, or NPU, and the app can optionally reach a local-network server like Ollama or LM Studio, but it needs no cloud account or API key to work.
Beyond Standard Local LLM Chatbots
Most phone-based local LLM apps handle one modality — you get a GGUF chatbot and nothing else, so image generation, vision, and speech-to-text each need a separate app, a separate model download, and their own slice of RAM. Off Grid AI keeps chat, on-device Stable Diffusion, vision models, and Whisper transcription inside one app, routed through the same memory manager, so switching from a text answer to a camera-based question doesn't mean closing one app and opening another. That shared memory manager also means you feel the RAM pressure directly — the app tells you what's resident and what it's costing you before something silently fails.
Core Capabilities of Off Grid AI
- ✓Text generation on any GGUF model — Qwen 3, Llama 3.2, Gemma 3, Phi-4, or your own .gguf file — with streaming output, a thinking mode, and markdown rendering at 15-30 tok/s on flagship phones.
- ✓Hardware-aware acceleration: Adreno GPUs run 20-40 tok/s via OpenCL on a Snapdragon 8 Gen 2+ versus 15-30 on CPU, Apple Silicon uses Metal, and the app auto-picks the fastest backend it detects (overridable in Settings).
- ✓Experimental Hexagon NPU support, badged per-model so you know before a 4GB download whether your device and quant combination can actually use it.
- ✓On-device Stable Diffusion image generation across 20+ models (Absolute Reality, DreamShaper, Anything V5, and others), NPU-accelerated at 5-10s per image on Snapdragon and Core ML-accelerated on iOS.
- ✓Vision AI through SmolVLM, Qwen3-VL, or Gemma 3n — point the camera at a document or scene and ask a question, answered in about 7 seconds on flagship hardware.
- ✓On-device Whisper speech-to-text for hands-free input, plus a tool-calling loop (web search, calculator, date/time, device info, knowledge-base search) with runaway prevention.
- ✓A project knowledge base that chunks and embeds PDFs and text with a bundled MiniLM model, retrieves by cosine similarity, and stores everything locally in SQLite.
- ✓A model manager with Lean/Balanced/Aggressive loading policies, a per-model RAM cost readout, a manual Load Anyway override, and a notice (with reload option) when a model gets evicted mid-conversation.
Practical Applications and Workflows
- •Chat or code offline with Llama 3.2 or Qwen 3 while traveling or on a flight, with no network fallback needed.
- •Point the camera at a receipt, whiteboard, or document and have SmolVLM or Qwen3-VL read and describe it without a cloud OCR call.
- •Generate concept art or mockups locally with one of the 20+ bundled Stable Diffusion checkpoints instead of sending prompts to a hosted image API.
- •Upload a project's PDFs to the knowledge base and query them through the search_knowledge_base tool inside a normal chat.
- •Point the app at a home-network Ollama, LM Studio, or LocalAI instance to borrow a bigger model's compute while keeping the request off the public internet.
- •With Pro: dictate hands-free through voice mode and let a persona draft a calendar entry, email, or Linear/Notion/GitHub ticket for you to approve before it sends.
Getting Started with Off Grid AI
Grab the latest APK straight from GitHub Releases, or install the iOS App Store build, which also runs natively on Apple Silicon Macs via Mac Catalyst/iPad compatibility. To build from source: clone the repo, run npm install, then cd android && ./gradlew clean && cd .. && npm run android for Android, or cd ios && pod install && cd .. && npm run ios for iOS. The README lists Node.js 20+, JDK 17 with Android SDK 36, and Xcode 15+ as prerequisites, with a fuller build guide in docs/ARCHITECTURE.md.
Strengths
- ✓One app actually covers five modalities — text, image generation, vision, speech-to-text, and tool calling — instead of stitching together separate single-purpose local AI apps.
- ✓Detects and uses the acceleration your device actually has (OpenCL on Adreno GPUs, Metal on Apple Silicon) rather than defaulting everyone to CPU-only inference.
- ✓MIT-licensed with the source and build steps public; the repo carries 2,875 GitHub stars and 276 forks.
- ✓Test coverage spans all three platforms — Jest/RNTL, JUnit, XCTest — plus Maestro end-to-end flows, run through BrowserStack on every PR.
- ✓The free tier needs no account or API key, and per the README no data leaves the device, even for the knowledge base's embeddings, which are stored in local SQLite.
Understanding Current Limitations
- △The Hexagon NPU path is marked experimental in the README: it only accelerates Q4_0 and Q8_0 quantizations; K-quantizations are silently processed by the CPU, and certain model architectures reportedly come out garbled on it.
- △Phone RAM is a hard ceiling — even the Aggressive loading policy can refuse a model, forcing a manual Load Anyway override, and a model can still get evicted mid-conversation under memory pressure.
- △Voice talkback, custom personas, action-drafting through MCP servers, and phone-Mac sync all sit behind the paid Pro tier ($69 one-time or $49/year); the free tier's voice feature is transcription only.
- △Phone-to-Mac sync was, per the README, still landing rather than shipped, so early Pro buyers don't get that cross-device continuity yet.
Other On-Device AI Options
Who Benefits from Off Grid AI?
Off Grid AI suits developers and privacy-conscious users who want a full offline AI toolkit — chat, image generation, vision, transcription — on a phone or Mac without a subscription to a cloud API, plus anyone testing GGUF models against real mobile GPU/NPU hardware. It's a poor match for anyone who needs a 70B-class model (phone RAM won't hold it) or team-shared conversation history, since sync is Mac-and-phone-only and, per the README, was still on the way rather than shipped.
Common Questions About Off Grid AI
Off Grid AI's core app — chat, image generation, vision, voice transcription, and tool calling — is free. Off Grid AI Pro adds voice talkback, personas, and action drafting for $69 once or $49 a year.
Off Grid AI runs on Android and iOS phones, and on Apple Silicon Macs through the iOS App Store build via Mac Catalyst/iPad compatibility.
Off Grid AI's chat, image generation, vision, and voice transcription all run locally, so no internet connection is required. It only reaches the network if you choose to connect to a remote server like Ollama.
Off Grid AI runs any GGUF text model, including Qwen 3, Llama 3.2, Gemma 3, and Phi-4, plus SmolVLM, Qwen3-VL, and Gemma 3n for vision, Whisper for speech, and 20+ Stable Diffusion checkpoints for images.
Off Grid AI Pro, a paid tier priced at $69 for life or $49 per year, brings several features: it enables on-device Kokoro text-to-speech when using voice mode, supports custom personas that possess persistent memory, and allows for draft-then-approve actions over Calendar, email, and MCP servers such as Linear, Notion, and GitHub.
Off Grid AI runs inference, transcription, and image generation entirely on-device, and per the README no data leaves the device and no account or API key is required; knowledge-base embeddings stay local in SQLite too.