Aggressive mixed-precision GGUF quants — shrink models to a fraction of their size, keep their edge and speed. Runs on llama.cpp (AMD/Apple too).