The purpose of this study was to explore the degree of grain size of the attributes and the sample sizes that can support accurate parameter recovery with the General Diagnostic Model (GDM) for a large-scale international assessment. In this resampling study, bootstrap samples were obtained from the 2003 Grade 8 TIMSS in Mathematics at varying sample sizes from 500 to 4000 and grain sizes of the attributes from a unidimensional model to one with ten attributes. The results showed that the eight-attribute model was the one most consistently identified as best fitting. Parameter estimation for more than ten attributes and samples less than 500 failed. Furthermore, the precision of item parameter recovery decreased as the number of attributes measured by an item increased and sample size decreased. On the other hand, the distributions of latent classes were relatively stable across all models and sample sizes.