Raspberry Pi 5 runs 122B LLMs via SSD swap

This technical stress test demonstrates that "local LLM" is a spectrum from interactive edge applications to slow-motion academic curiosities. Sub-4B parameter models like Qwen 0.8B and 2B maintain usable speeds of 5–11 t/s, while SSD-backed swap allows loading massive 122B models that would otherwise fail. Thermal data confirms a "wait-bound" cooling effect, as CPU temperatures drop while idling for I/O during heavy swap operations.