Considering the accuracy and diverse balance problem of ensemble learning, a selective ensemble extreme learning machine with kernels based on mutual information is proposed. The extreme learning machine with kernels, which has characteristics of simple structure, fast training and good generalization, is chosen as the base learner. The correlation and redundancy criterion are introduced to describe the accuracy and diverse, and the selective ensemble problem is transformed as a variable problem. Then, the maximum correlation minimum redundancy criterion based on mutual information is used to select the extreme learning machine with kernels. The simulation results based on UCI benchmark regression and classification data show the advantages of the proposed algorithm.