Airux Pocket AI
An open-source local LLM chat app for Android that runs GGUF models with llama.cpp, processes photos on-device, and offers optional DuckDuckGo search.
Airux Pocket AI runs text and vision models directly on arm64 Android phones running Android 8 or later. It's for people who want private AI chat on their phone, including offline conversations and photo analysis without cloud inference. You don't need an account. Chat history and downloaded models stay in the app's private storage.
The app uses llama.cpp to run GGUF models, with a catalog that includes Qwen3, Llama 3.2, Gemma 3 and SmolVLM2. Model choices depend on the phone's RAM: smaller Llama and SmolVLM2 models start at 4 GB, while larger choices require 6 or 8 GB. The app adapts prompts and available capabilities to each model rather than treating every model alike.
Vision models can work with attached photos locally, using the same chat session for images and text. Conversations retain their own context, and a coding mode supports continuing longer code responses. Model-specific prompts shape identity and coding behavior.
Models download from Hugging Face, so getting them requires an internet connection. Once downloaded, they support offline chat. Optional web search sends queries to DuckDuckGo when a supported model requests it; Qwen3 1.7B and Llama 3.2 3B support this without API keys or a cloud service coordinating the chat.
The app is open source under the MIT license. Downloaded models have their own upstream licenses.
