The paper introduces a novel implementation of the harmonic-balance (HB) technique, explicitly devised for the efficient simulation of large-size nonlinear microwave circuits. The method is very powerful and general, and can be applied both in a piecewise and in a nodal harmonic-balance environment. The problem unknowns are subdivided into a master and a slave set, and the HB system is solved hierarchically by a two-level Newton algorithm. CPU time and memory savings are dramatic when the master set is small and the slave system may be decomposed into several smaller subsystems. Families of circuits for which this situation occurs in practice are discussed and are shown to cover most situations of practical interest.