In a MoE model with experts_per_token = 2 and each expert having 7B params, after picking the experts it should run as fast as the slowest 7B expert, not a comparable 14B model.