Since Strix Halo is a rather expensive 256-bit APU/chip (equivalent to a quad-channel PC/workstation, but it's 8000 MT/s), using anything less than the 128 GB RAM it supports is a waste (one could say a waste of silicon). Even some lightweight-ish MoE LLMs easily require all the 128 GB RAM they can get: huggingface.co/unsloth/GLM-4.5-Air-GGUF.