Accurate trajectory prediction is essential for safe and efficient motion planning in autonomous driving. Most existing models use Transformer- or graph-based architectures to model interactions between agents in traffic environments. However, their complexity often limits real-time prediction in interaction-rich scenarios. Mamba, a recent advancement in state space models, leverages linear time complexity and hardware-aware design to improve inference efficiency, showing strong performance in sequence modeling tasks. In this paper, we introduce a novel and efficient trajectory prediction architecture built on Mamba. Specifically, we propose a new encoding module that applies cross-attention at each timestep to separately encode the static and dynamic environments surrounding the target vehicle. These features are then fused and fed into a unidirectional Mamba model to extract rich temporal information. Additionally, we adopt a reparameterization technique during decoding to generate multimodal future trajectories. Unlike traditional reparameterization methods that rely on random sampling, our approach uses parameter-guided sample prediction, reducing the likelihood of generating unreliable trajectories. Experiments on the INTERACTION and highD datasets demonstrate that our approach achieves state-of-the-art performance.