Multiple Input Multiple Output (MIMO) technology is widely applied in various wireless communication systems, significantly improving communication efficiency and reliability. Signal detection is critical for MIMO systems. However, with the increasing integration of deep learning into MIMO signal detection algorithms, challenges such as high complexity and limited interpretability have emerged. To address this, this paper proposes a model driven trainable approximate message passing (AMP) algorithm that combines the iterative process of AMP with deep learning techniques. By introducing trainable parameters and optimizing them through training, and incorporating an attention mechanism to enhance channel feature extraction, the detection accuracy is improved, and the algorithm’s generalization capability is enhanced. Simulation results demonstrate that AMP Attention Net achieves lower bit error rates compared to traditional detection algorithms. Furthermore, the proposed algorithm exhibits robust performance under different configurations of transmitting and receiving antennas.