Little sparse on detail, I regularly run LLMs on 5 year old CPUs so no problem there, I wonder how the approach compares in memory requirements to existing quantization methods.