Definitely experiment with quantization. In my experience, a rough rule is to just use the biggest model you can at the smallest quantization.