Offgrid

Offgrid runs a large language model directly on your phone's processor. There is no server doing the thinking, no account to create, and nothing you say to a model is sent anywhere. Once a model is downloaded, chatting works with the radios off. WHY THIS IS DIFFERENT Most "AI" apps are a text box wired to somebody else's server. Your conversation leaves your device, gets logged, and stops working when you lose signal. Offgrid does the inference itself, using llama.cpp compiled for your phone's CPU. Put it in airplane mode and it keeps answering. PICK A MODEL THAT ACTUALLY FITS Search every public GGUF model on Hugging Face from inside the app. Each quant is checked against your phone's memory before you download it, so you find out that a 4 GB model will not run on a 4 GB phone before spending the bandwidth, not after. There is a short curated list too, if you would rather not choose. Offgrid also refuses to load a model that cannot work. It reads the model's real geometry, works out what the weights and the context window will actually cost, and offers a context size that fits instead of letting the app get killed. PROFILES Tell it what you are using it for and the system prompt, sampling settings and model ordering follow. Everyday assistant, off-grid reference, coding, writing, studying, or ask-anything. BUILT FOR REAL PHONES - Foldables and tablets, with a resizable sidebar - Light and dark themes, reduce-motion and haptics settings - Download manager that pauses, resumes and survives leaving the app - Automatic thread tuning for your specific CPU - Share or select text anywhere on the phone and send it straight to the model FOR DEVELOPERS An optional local server exposes the loaded model over OpenAI-compatible endpoints, so your own scripts and apps can use it. Off by default, bound to localhost unless you deliberately change it. PRIVACY, PLAINLY No account. No analytics. No crash reporting. No server of ours at all. Conversations live in this app's private storage and are deleted when you uninstall it. The only time anything leaves your device is when you search for or download a model, because that is when the app talks to Hugging Face. A NOTE ON MODELS Models are third-party files made by other people. What a model says is a property of that model, not of this app. The built-in list contains only models published with their original safety training intact. You can report any reply from inside the app. REQUIREMENTS A 64-bit ARM phone from roughly 2018 onward (Cortex-A75/A55 or newer). Offgrid checks before it runs and tells you if your processor cannot support it rather than crashing. 4 GB of RAM is comfortable for small models; more is better.