The paper introduces PivotMesh, a generic and scalable framework for generating compact and detailed 3D meshes. Unlike existing methods that often produce dense and over-smoothed meshes, PivotMesh aims to generate more human-crafted, compact meshes. The framework uses a transformer-based auto-encoder to encode meshes into discrete tokens and decode them hierarchically from face to vertex levels. It also employs pivot vertices as coarse representations to guide the mesh generation process, improving controllability and scalability. PivotMesh is evaluated on small datasets like Shapenet and large-scale datasets like Objaverse and Objaverse-xl, demonstrating superior performance in generating high-quality, compact meshes across various categories. The method's effectiveness is validated through extensive experiments, showing its potential for native mesh modeling and downstream applications such as mesh variation and refinement.The paper introduces PivotMesh, a generic and scalable framework for generating compact and detailed 3D meshes. Unlike existing methods that often produce dense and over-smoothed meshes, PivotMesh aims to generate more human-crafted, compact meshes. The framework uses a transformer-based auto-encoder to encode meshes into discrete tokens and decode them hierarchically from face to vertex levels. It also employs pivot vertices as coarse representations to guide the mesh generation process, improving controllability and scalability. PivotMesh is evaluated on small datasets like Shapenet and large-scale datasets like Objaverse and Objaverse-xl, demonstrating superior performance in generating high-quality, compact meshes across various categories. The method's effectiveness is validated through extensive experiments, showing its potential for native mesh modeling and downstream applications such as mesh variation and refinement.