Run Gemma 2 offline on Android, with iPhone setup notes

This tutorial covers a phone app for running a local LLM on Android, with separate GitHub setup instructions for iPhone and iPad. The presenter describes the app as open source and demonstrates replies after disabling Wi-Fi and mobile data. These demonstrations support the offline workflow shown; the broader claims about complete privacy and unrestricted answers are the presenter's claims.

Android installation starts with the APK linked on GitHub and the recommended ARM64 build. After obtaining the app and model files, the walkthrough imports a GGUF file, copies it into the app's storage, then loads the model into RAM. The presenter recommends Gemma 2 2B for phones with 4 GB of RAM, describing the download as roughly 1.6 GB and the imported file as about 1.7 GB. The presenter estimates a 10–20-second wait for the first reply in this setup. He also reports trying Gemma 4, which he says runs more slowly than Gemma 2.

The app includes chat history, temperature and system prompt settings, plus a local API server with OpenAI-style endpoints. Users can add a model by pasting its GGUF URL from Hugging Face, then download and load it. Model choice must match the phone's specifications. iOS requires the linked guide rather than Android's APK installation process. A comparison with ChatGPT focuses on one prompt and the presenter's assessment of the responses, rather than a general capability benchmark.