In this paper, we propose PlanFormer, a transformer-based motion planning model tailored for autonomous driving. Our approach emphasizes improved contextual understanding while maintaining computational efficiency, setting it apart from prior methods that often rely on complex and resource-intensive encoder architectures. By simplifying the encoder and allowing self-attention mechanisms to directly operate on structured motion planning inputs–including historical agent trajectories, dynamic objects, and map information–we enable the model to more effectively capture spatiotemporal dependencies in diverse driving scenarios. To further enhance model capability, we incorporate three recent architectural innovations: RMSNorm for improved training stability, SwiGLU activation for enhanced expressiveness, and rotary positional embeddings to preserve relative spatial and temporal relationships in the input data. Together, these enhancements contribute to more accurate and robust planning decisions under varied real-world conditions. We validate our approach on the large-scale nuPlan benchmark, which provides closed-loop simulation environments and diverse driving scenarios. Our evaluations focus particularly on the reactive closed-loop setting, where agents interact realistically and the planner’s performance most closely reflects real-world deployment challenges. Experimental results show that PlanFormer achieves state-of-the-art performance among pure learning-based methods, excelling in both decision-making efficiency and trajectory accuracy. By striking a balance between architectural simplicity and deep contextual reasoning, our method offers a scalable and effective solution for next-generation autonomous planning systems.