PointNet++ is one of the most influential neural architectures for pointcloud understanding. Although the accuracy of PointNet++ has been largelysurpassed by recent networks such as PointMLP and Point Transformer, we findthat a large portion of the performance gain is due to improved trainingstrategies, i.e. data augmentation and optimization techniques, and increasedmodel sizes rather than architectural innovations. Thus, the full potential ofPointNet++ has yet to be explored. In this work, we revisit the classicalPointNet++ through a systematic study of model training and scaling strategies,and offer two major contributions. First, we propose a set of improved trainingstrategies that significantly improve PointNet++ performance. For example, weshow that, without any change in architecture, the overall accuracy (OA) ofPointNet++ on ScanObjectNN object classification can be raised from 77.9% to86.1%, even outperforming state-of-the-art PointMLP. Second, we introduce aninverted residual bottleneck design and separable MLPs into PointNet++ toenable efficient and effective model scaling and propose PointNeXt, the nextversion of PointNets. PointNeXt can be flexibly scaled up and outperformsstate-of-the-art methods on both 3D classification and segmentation tasks. Forclassification, PointNeXt reaches an overall accuracy of 87.7 on ScanObjectNN,surpassing PointMLP by 2.3%, while being 10x faster in inference. For semanticsegmentation, PointNeXt establishes a new state-of-the-art performance with74.9% mean IoU on S3DIS (6-fold cross-validation), being superior to the recentPoint Transformer. The code and models are available athttps://github.com/guochengqian/pointnext.