The multichannel blind least mean square (MBLMS) algorithm is proposed to implement mode demultiplexing for a 2×2 mode division multiplexing （MDM） system. The principle of MBLMS algorithm based on the non-Gaussian maximization is described. The demultiplexing performance of this algorithm is analyzed, and this algorithm is compared with the least mean square (LMS) algorithm based on data assistance. The simulation results show that the MBLMS algorithm can realize blind separation of the convolved mixed signals without data assistance. Its demultiplexing performance is comparable with that of the LMS algorithm, and its convergence rate is improved by 33.3% compared with that of LMS algorithm based on data assistance.