Comment by bigyabai

The model is 80b parameters, but only 3b are activated during inference. I'm running the old 2507 Qwen3 30B model on my 8gb Nvidia card and get very usable performance.