Multivariate time series (MTS) forecasting provides significant benefits for various real-world applications, yet it imposes essential demands on the models’ capacity to effectively capture both temporal and variable relationships. Recently, DL-based models have significantly improved MTS forecasting performance by leveraging their ability to model complex dependencies. However, challenges remain due to the weak semantic information of individual time steps, which limits the comprehensive capture of temporal and variable interactions. To address the limitations of existing models, we propose BiG-Mamba, a novel framework designed to simultaneously capture dynamic temporal-level and variable-level dependencies. First, our approach integrates a carefully designed frequency-domain decomposition, which divides the time series into sub-series at multiple scales. Second, BiG-Mamba utilizes a synergistic combination of Graph Convolutional Network (GCN) and Mamba, effectively leveraging their strengths to capture the temporal patterns across multiple time scales. Third, BiG-Mamba leverages an adaptive GCN to dynamically adjust and capture critical associations between key variables. Extensive experiments on seven benchmark datasets demonstrate the effectiveness of BiG-Mamba in capturing intricate time series and its superiority over state-of-the-art methods in delivering enhanced forecasting accuracy.