Data-driven deep learning (DL) has delivered impressive performance gains in physical layer communications in recent years. However, the black box nature of these models makes them difficult to interpret and requires massive training datasets, which hinders their practical deployment and future development. A state-of-the-art architecture, known as model-driven DL, offers tremendous potential for addressing these issues by properly relating deep neural networks to models and algorithms with strong theoretical foundations in signal processing. However, there is no instructive architecture reported for model-driven DL-based multiple-input and multiple-output (MIMO) transceiver design. This article presents the various MIMO transceiver modules as generalized linear models (GLMs), and reviews model-driven DL-based GLMs for these modules to tackle this problem. The characteristics of these methods are analyzed, a universal model-driven DL framework for MIMO transceivers is developed, and performance evaluation in practical scenarios is conducted. Finally, future directions for the universal architecture from the perspective of generalizability and computational complexity are presented.