Recent advances in deep learning have substantially enhanced gait recognition performance. However, the increasing parameter size of large-scale models has become a major limitation, hindering further progress. Since binary silhouette sequences used in gait recognition are typically of variable length and contain considerable redundancy, it is essential to develop efficient and lightweight methods for their effective processing. To this end, we propose SMEGNet, a lightweight MLP-based framework for variable-length gait sequences. Specifically, SMEGNet introduces a proposed Fixed-Length Sequence Truncation method to standardize input sequence lengths. Furthermore, it incorporates a Cross-Dimensional Attention module to capture comprehensive spatial, channel, and temporal dependencies, as well as a Block-wise MLP module to enhance spatiotemporal feature modeling through structured tensor partitioning. Extensive experimental results demonstrate that the proposed method achieves competitive recognition accuracy while significantly reducing the number of parameters compared to existing methods.