This is hard a-priori, but fairly easy post-facto. Model distillation isn't a common practice yet, but it has already been demonstrated to be quite effective for specific use cases.