Apple Hasn’t Put an LLM on Your Watch. So I Did.

A developer successfully ran a real language model, generating text at 24 tokens per second, completely offline on an Apple Watch Series 6, a six-year-old device. This was achieved despite watchOS not being officially supported by llama.cpp. The project involved overcoming architectural assumptions, addressing the 2GB memory ceiling, and using specific build flags. The full Xcode project, including two models and a tool-calling setup, is available on GitHub.