Or distilled models, or just slightly smaller models but same architecture. Lots of options, all of them conveniently fitting inside "optimizing inferencing".