Experimental. A small AI model (llama.cpp via wllama) runs on this device, with no internet after the first use. The model is downloaded once from Hugging Face and kept in the browser. Sizes: Qwen 0.5B 0.4 GB (fastest), Llama 3.2 1B 0.8 GB, Qwen 2.5 1.5B 1.1 GB. Small models only: good for short chats, summaries and fixing text, but they make mistakes and can be slow on phones and tablets. In the Windows app every model runs on the PC with native llama.cpp, from Qwen 0.5B to Llama 3.3 70B (42.5 GB), on the graphics card or, without one, on the processor; the model size is the free memory it needs. The + button adds files (text, PDF, Word, PowerPoint, Excel, OpenDocument, CSV, code) for the AI to read; in the Windows app it also takes images (Qwen VL models) and reads web pages whose address you write in the message. Nothing you type leaves the device. wllama and llama.cpp are free software (MIT).