Electroencephalogram (EEG) motor imagery decoding has become an essential non-invasive technology for brain–computer interfaces. However, simultaneously capturing multi-scale temporal patterns, integrating cross-electrode spatial information, and fusing heterogeneous pooling statistics remains challenging. To address these challenges, we introduce a novel framework termed Multi-Scale Hybrid-Pooling Attention Transformer Network (MaHPA-Net) for MI-EEG decoding. The proposed MaHPA-Net integrates parallel convolutional operations with multi-scale receptive fields, a hybrid-pooling–driven channel attention strategy, and a unified spatial fusion mechanism. By incorporating a CLS Token with positional encoding into a Transformer encoder to model global temporal dependencies, MaHPA-Net realizes end-to-end spatio-temporal representation learning without relying on hand-crafted adjacency matrices. Experimental results on the BCIC-IV-2a and IV-2b datasets demonstrate the effectiveness of the proposed network, yielding average accuracies of 81.52% and 88.89%, along with high Cohen’s Kappa, F1-score, and Recall, which confirms its robust discriminative capability and strong generalization in MI-EEG classification.