I read about its format being tough to compress further. What should we expect in terms of creating smaller models to run locally?