Quantization the better approach in most cases, unless you want to for instance create hybrid models ie. distilling from here and there.