Expert parameters that are active for every token in an MoE model, providing common processing before token-specific expert routing.