<p>3D point clouds are known for the complexity of their processing due to their irregular and disordered nature. These characteristics make it particularly difficult to extract rich features relevant to 3D applications. In this context, recent advances have shown that it is possible to use transformers to process 3D point clouds. However, the accuracy of these models is moderate compared to other approaches. To overcome this limitation, we present a Multi-Stage Convolutional Transformer for Point Clouds, which is applied to the classification of 3D point clouds. Our approach adopts a multi-stage transformer encoder that merges feature vectors from multiple levels. The proposed architecture integrates a Convolutional TransitionDown module and an Adaptive Conv-Attention module, exploiting the properties of convolutional projection. It also relies on an aggregation module to facilitate the fusion of multi-level features. The methodology used combines convolutions and linear transformations in the embedding layer as well as relative position encoding and vector attention. Simulation results on the datasets ModelNet40 and ScanObjectNN show that our approach demonstrates good performance compared to previous methods.</p>