Gliomas are among the most common malignant brain tumors in adults and present significant challenges in neuro-oncology due to their heterogeneity and complex biological behavior. Accurate segmentation of gliomas using multimodal magnetic resonance imaging is essential for diagnosis and treatment planning, but it remains computationally demanding. U-Net-based models, such as Attention U-Net, UNETR, and Swin UNETR, have demonstrated strong performance in brain tumor segmentation but face inherent limitations in capturing complex nonlinear patterns. Kolmogorov-Arnold Networks (KANs) introduced learnable univariate spline functions to enhance modeling capabilities; however, their computational inefficiency poses challenges for practical deployment. In response, we propose µPUA-Net, which integrates PowerMLP layers into the UKAN-EP architecture, replacing KAN layers to improve computational efficiency while maintaining strong modeling capacity. Our approach achieves comparable accuracy to UKAN-EP with a 36.3% reduction in parameters and improves efficiency over UKAN-EP, with the lowest 3% error in accuracy compared to nn-UNet. Furthermore, our model achieves a remarkable 1/2159 reduction in inference time compared to nn-UNet, highlighting its potential as a practical and effective solution for glioma segmentation tasks.