To enhance the adaptability and safety of autonomous vehicles in complex environments, accurate environmental understanding and prediction are paramount, motivating research into world models. Point cloud-based prediction world models have garnered significant attention due to their superior three-dimensional environment representation capabilities. However, in unstructured off-road scenarios, existing methods still face challenges in achieving accurate predictions of terrain edge features and ensuring the accuracy of continuous state prediction. To solve these problems, we propose the self-supervised point cloud prediction world model WildPCP-Mamba. This model, based on the Mamba architecture, constructs a bidirectional spatio-temporal modeling network capable of effectively learning spatiotemporal features from point clouds. Additionally, the integrated Terrain Edge Feature Enhancement Module (TEEM) employs attention mechanisms and multi-level feature fusion to significantly improve edge feature prediction performance. Experiment results on the Rellis-3D dataset demonstrate WildPCP-Mamba’s superior point cloud prediction capabilities, with an average Chamfer distance reduction of 43%. The model achieves precise predictions up to 5 frames while maintaining a Chamfer distance below 0.2.