“A limitation of current Mixture-of-Experts models is their regular structure, where each expert component is roughly the same size.”