BrepGen is a diffusion-based generative model that directly outputs Boundary Representation (B-rep) CAD models. It represents B-rep models as a structured latent geometry in a hierarchical tree, where each element (face, edge, vertex) is encoded as a node. The model uses a Transformer-based diffusion approach to sequentially denoise node features, recovering B-rep topology information through node duplication. BrepGen generates complex B-reps with free-form and doubly-curved surfaces, surpassing existing methods on various benchmarks. It introduces a new Furniture B-rep Dataset with high-quality B-reps of indoor objects across 10 categories. The model also supports CAD autocomplete and design interpolation. BrepGen's structured latent geometry representation unifies geometry and topology information in a tree format, enabling effective training and inference. The model outperforms existing methods in generating watertight solids and handles complex CAD models with multiple bodies. However, it has limitations, such as not guaranteeing watertight solids and potential issues with self-intersections or noise in decoded points. BrepGen demonstrates state-of-the-art results in B-rep generation and moves towards an automatic system capable of directly generating B-reps to reduce manual labor in CAD design.BrepGen is a diffusion-based generative model that directly outputs Boundary Representation (B-rep) CAD models. It represents B-rep models as a structured latent geometry in a hierarchical tree, where each element (face, edge, vertex) is encoded as a node. The model uses a Transformer-based diffusion approach to sequentially denoise node features, recovering B-rep topology information through node duplication. BrepGen generates complex B-reps with free-form and doubly-curved surfaces, surpassing existing methods on various benchmarks. It introduces a new Furniture B-rep Dataset with high-quality B-reps of indoor objects across 10 categories. The model also supports CAD autocomplete and design interpolation. BrepGen's structured latent geometry representation unifies geometry and topology information in a tree format, enabling effective training and inference. The model outperforms existing methods in generating watertight solids and handles complex CAD models with multiple bodies. However, it has limitations, such as not guaranteeing watertight solids and potential issues with self-intersections or noise in decoded points. BrepGen demonstrates state-of-the-art results in B-rep generation and moves towards an automatic system capable of directly generating B-reps to reduce manual labor in CAD design.