I’ve got an M4 Mac mini with 32 GB of unified memory, so I can run MoE models in the <40B parameter range quite comfortably. I use LM Studio as the backend with Hermes as the frontend. Works great for everyday chat, web search, and RAG.