Accurate time series forecasting is crucial across various domains. Multivariate forecasting requires models that capture complex temporal dependencies and variable interactions. In recent years, Transformer-based models have achieved strong efficacy by specializing in capturing long-range temporal patterns through self-attentive mechanisms. However, it suffers from quadratic complexity and isotropic token processing, limiting scalability for high-dimensional data. In this paper, we propose BiMa-Former, a novel hybrid architecture that alternately stacks temporal Transformer blocks and multivariate Mamba blocks to model both temporal and variable dimensions. We design a dual-token embedding layer in which temporal tokens are independently generated via convolution, while variate tokens are produced through a linear mapping, thereby decoupling temporal and variable features. The temporal blocks use Rotary Positional Encoding (RoPE) attention and gated residual connections to fuse multi-scale temporal patterns, while the variable blocks employ bidirectional scanning to capture full-variable dependencies with linear complexity, preserving the benefits of State Space Models (SSMs). Finally, the generation of forecast outcomes through a linear mapping layer. Extensive experiments and ablation studies on three benchmark datasets demonstrate that BiMa-Former flexibly accommodates diverse data characteristics, establishing a new paradigm for efficient and scalable hybrid modeling in multivariate time series forecasting.