<p>Multivariate time series forecasting requires effective modeling of temporal dependencies and evolving relationships among variables. Although recent methods have achieved strong predictive performance, many of them rely on computationally intensive architectures, which limits their applicability in resource-constrained scenarios such as IoT monitoring, edge computing, and real-time decision-making. This motivates the development of forecasting models that can maintain competitive accuracy while reducing computational and memory overhead. In this paper, we propose MFSNet, a lightweight forecasting framework that combines hidden-space multi-scale representation learning with differentiable frequency-domain suppression. Specifically, MFSNet first adopts a lightweight multi-scale patch embedding strategy, where parallel multi-granularity max-pooling operations are performed in the hidden space to capture temporal patterns at different scales with limited additional cost. This design avoids the substantial overhead typically introduced by explicit multi-scale stacking or complex routing schemes. The resulting fused representations are then passed to an MLP-guided frequency suppression (MFS) module, which learns differentiable spectral modulation coefficients to adaptively suppress redundant frequency components. In this way, the proposed module enhances the temporal modeling capability of MLP-based architectures, particularly for global periodic patterns, while preserving structural simplicity. A lightweight prediction head based on flattening and linear projection is finally used to generate future forecasts. Experimental results on eight real-world benchmark datasets show that MFSNet achieves competitive forecasting performance across diverse settings. In addition, it demonstrates favorable inference efficiency and memory usage under standardized evaluation, indicating its potential as a practical lightweight solution for multivariate time series forecasting.</p>