Component mode synthesis is commonly used to simulate the structural behavior of complex systems. Among other component mode synthesis techniques, the Craig–Bampton method stands out for its popularity. However, for finely meshed systems featuring many components, the size of the resulting assembled system is dominated by the interface degrees of freedom. The system-level interface reduction technique aims at reducing the size of the assembled reduced model by extracting a few dominant interface modes. If the size of the interface degrees of freedom is large, the resulting problem is almost as computationally expensive as the one associated to the full model. Conversely, the local-level interface reduction technique reduces the interface of each substructure before assembly. In this case, the computational effort associated to the local eigenvalue problem is moderate, but issues arise when enforcing compatibility between interfaces. In this paper, the computational effort related to the interface reduction is significantly reduced by performing two variants of the multilevel Craig–Bampton reduction when the subsystems are assembled in subsets. This procedure localizes the interface reduction by applying a multilevel static condensation and eigenvalue analysis on each subset in parallel. The different interface reduction techniques are assessed on large-size realistic examples.