These models work well with http://github.com/Atomic-Germ/Guanaco -- a disk-streaming MoE focused llama.cpp built with patches to upstream main.