It is hard for traditional precoding algorithms to balance the high performance and computational complexity. Therefore, a low-complexity MLP-mixer-based precoding network (MMPNet) is proposed with unsupervised training in this chapter. The feasibility of using the compressed channel correlation matrix as input of the DL network is first proved. Subsequently, a low-complexity feature extraction module is designed to fit for the input structure, and the recovery module of the WMMSE algorithm is also improved. Finally, the scalability of the proposed model versus the varying numbers of UEs is discussed.