Accurate, real-time estimation of lower limb joint moments is critical for advancing biomechanics and rehabilitation technology. This study benchmarks a state-of-the-art Temporal Convolutional Network (TCN) against two advanced time series models, gMLP and TimeGPT, for joint moment estimation. To simulate a reduced sensor suite, we used only bilateral hip angle data to simultaneously estimate joint moments in both limbs. We collected experimental data across three distinct gait conditions, stand-to-walk transition, level-ground walking, and transitional speed walking, to evaluate model performance. Beyond estimation accuracy, we assessed computational efficiency by comparing floating-point operations (FLOPs) and model parameter counts to determine real-time deployment feasibility. Our results show that while TCN maintains high accuracy across all conditions, TimeGPT offers competitive performance with significantly lower computational complexity. Meanwhile, gMLP struggles to model the complexities of functional gait conditions. This comparative analysis provides a clear guide for selecting the optimal model for biomechanical applications, balancing the need for precision with computational scalability.