The rapid evolution of laser radar technology has enabled the widespread acquisition of three-dimensional point cloud data, posing challenges in real-time analysis and efficient information extraction. The focus of this research is to solve a key problem in 3D point cloud data processing: to achieve rapid and efficient analysis of point cloud classification and semantic segmentation tasks. We propose the AnisoVector network, which divides local point cloud processing into two distinct stages: point position processing and point feature processing, thereby significantly enhancing the model’s inference speed. By introducing anisotropic functions, our approach enables Multilayer Perceptrons (MLPs) to differentially process each local point cloud, thus augmenting the model’s capacity for fine-grained point cloud analysis (Separable Anisotropic - SAM). Furthermore, vectors with size and directional properties enhance the flexibility of point cloud representation, enabling the network to handle diverse point cloud shapes more effectively and enhance the ability to extract local point cloud features (Point Vector - PVM). By employing a group vector attention mechanism, efficient capture of feature information within local point clouds has been achieved, thereby enhancing the efficiency and generalization performance of the model (GroupFormer - GFM). Experimental results across ModelNet40, ShapeNetPart, and S3DIS datasets demonstrate superior performance in tasks like classification and segmentation.