It is demonstrated that the crystallographic models of macromolecules may appear to diverge upon extended refinement against experimental data. Two regimes are identified for this phenomenon. Firstly, at higher resolution the apparent instability of the resulting models is shown to originate from the relatively small fraction of disordered atoms present in the initial model. Secondly, at lower resolution additional refinement instability may arise from insufficiently strong geometry restraints. The convergence of crystallographic refinement is proposed as one of the possible criteria in selecting a specific refinement strategy and in model validation.