To accurately establish the soft sensor model of cement fineness, the selection of input variables is easily influenced by time delay. Thus, this paper proposes a soft sensor modeling method based on mutual information and least square support vector machine (MILSSVM). In the proposed method, the correlation of variables is represented by mutual information, which can be used to determine the time delay of each auxiliary variable. Furthermore, the twoway selection algorithm is proposed to obtain input variables, and a soft sensor model of cement fineness is built based on least squares support vector machine with these selected input variables. Finally, the proposed model is trained with actual operational data of a cement plant. The experimental results show that the model can achieve high precision and generalization.