Multivariate time-series (MTS) forecasting has wide-ranging applications in areas such as finance and energy. For multivariate time-series data, it is crucial to model both temporal dependencies and inter-variable correlations simultaneously. Transformer-based models have shown great promise for capturing long-range temporal dependencies and inter-variable correlations. However, existing approaches often prioritize one aspect over the other, leading to unsatisfactory results. To address this gap, we propose a method called MVTformer. The method employs improved point-token and sequence-token strategies, each integrated with self-attention, to capture temporal dependencies and inter-variable correlations, respectively. To obtain a unified and rich representation, we introduce a cross-attention mechanism to align the temporal and variable branches, and a gating mechanism to achieve the final fusion. To further refine these representations, we employ a feedforward network to model nonlinear relationships, thereby enhancing the model’s generalization capability. We have conducted extensive experiments to validate our method; across seven publicly available datasets, it achieved the best results on more than 65% of tasks and markedly outperformed competing methods.