Empirical studies that focus on the impact of three-dimensional (3D) visualizations on learning are to date rare and inconsistent. According to the ability-as-enhancer hypothesis, high spatial ability learners should benefit particularly as they have enough cognitive capacity left for mental model construction. In contrast, the ability-as-compensator hypothesis proposes that low spatial ability learners should gain particular benefit from explicit graphical representations as they have difficulty mentally constructing their own visualizations. This study examines the impact that interactive 3D models implemented within a hypermedia-learning environment have on understanding of cell biology. Test scores in a subsequent knowledge acquisition test demonstrated a significant interaction term between students' spatial ability and presence/absence of 3D models. Only students with high spatial ability benefited from the presence of 3D models, while low spatial ability students got fewer points when learning this way. When using 3D models, high spatial ability students perceived their cognitive load to be low whereas the opposite was true for low spatial ability students. The data suggest that students with low spatial ability became cognitively overloaded by the presence of 3D models, while high spatial ability students benefited from them as their total cognitive load remained within working memory limits.