A model that I can run on a single consumer GPU is as intelligent as gpt-5.6-luna? That's insane.