I have had a difficult time with running 120b models on my 128gb setup, especially with any larger context size. The 6bit of Qwen 3.5 is already just over 100gb, and when you go down to 4bit it seems a bit lobotomized.