"The MOE architecture uses 20 times the parameters, is this comparison fair? Can it be compared with a single model that also uses 20 times the parameters?"