Are you offloading this to your gpu? I've always suck to GPTQ models because CPU quantuzations can never be as good.