I didn't know this was actually possible until today. Using https://github.com/antirez/ds4#running-models-larger-than-ram Antirez's specific engine + his specific ds4 gguf it literally just runs. You need to pass --ssd-streaming When running if you have <128gb I think. Seems 64gb and up is reasonable. I also passed: iogpu.wired_limit_mb=86016 To raise available metal allocation then you can patch the repo itself to increase cache safety which is .70 optionally to try and pu...