In the computational fluid dynamics study of hydrocyclones, the air-core dimension is key to predicting the mass split between the underflow and overflow. In turn, the mass split influences the prediction of the size classification curve. Three models, the renormalization group K-epsilon model, the Reynolds stress model, and the large-eddy simulation model, are compared for the predictions of air-core dimension, mass split, and axial and tangential velocities. The large-eddy simulation model, since it produces some detailed features of the turbulence, is clearly closer in predicting the experimental data than the other two. It is shown that particle tracking done with the velocity field obtained from the large-eddy simulation model accurately predicts the experimental size-classification curve. (c) 2005 Elsevier B.V. All rights reserved.