I've tested some local LLMs on prosumer hardware, here are some findings