Combining multilevel (ML) analysis and multidimensional item response theory (MIRT) provides a valuable method for analyzing data of educational assessments, where clustered data (e.g., students in classes) and multidimensional constructs frequently occur. It allows to model multiple ability dimensions while simultaneously taking the hierarchical structure into account. The dimensional structure of students’ foreign language competence within and between classrooms was investigated by applying a ML-MIRT measurement model to data of N = 9,410 students in 427 classes who had answered three different subtests of English as a foreign language. Results were compared to a MIRT model not taking into account the multilevel structure. A markedly more differentiated correlation structure is found within classrooms compared with the between-classroom level and compared with the model without multilevel structure. Results show that by modeling the latent multilevel structure, estimation and interpretation of ability profiles can be possible even with highly correlated ability dimensions