What you need to know
Offline AI on Android requires an app or installed web app, a compatible local runtime, and a model that is fully stored on the device. Download and activate everything while connected, then test a new prompt in airplane mode. Web search, model downloads, updates, cloud sync, and other online features will remain unavailable.
Prepare the complete local path
Installing an interface is only the first step. The application shell, model weights, tokenizer, runtime, and any document extraction components needed for your task must be present before connectivity disappears. Confirm the selected model is active, leave storage headroom for temporary files, and keep enough battery available for sustained computation.
A browser-based assistant can also work offline when its PWA shell and runtime are cached and its model remains in browser storage. WebGPU or WebAssembly can perform local computation, but neither guarantees that storage will persist or that a particular phone supports the model. Native and browser apps should both be tested in the context you plan to use.
Test the situation that matters
Do not rely on an old conversation that may already contain a response. Disable Wi-Fi and mobile data, open a fresh chat, and submit a prompt that requires new generation. If you need documents, import a representative file and confirm extraction, retrieval, and citations while still offline. Restart the app once to check that the model selection and local data survive a normal reopen.
When comparing offline apps, record the exact phone, app version, model artifact, quantisation, prompt, and context size. Check time to first visible output, cancellation, repeated use, heat, and recovery after an error. These observations are more useful than generic speed rankings because local performance depends on the whole device and runtime combination.