“A limitation of current Mixture-of-Experts models is their regular structure, where all experts are the same size and computational paths merge back together quickly.”