We present a new model for generating molecular data by combining discrete and continuous diffusion processes. Our model generates a comprehensive representation of molecules, including atom features, 2D discrete molecule structures, and 3D continuous molecule coordinates. The use of diffusion processes allows for capturing the probabilistic nature of molecular processes and the ability to explore the effect of different factors on molecular structures and properties. Additionally, we propose a novel graph transformer architecture to denoise the diffusion process. The transformer is equivariant to Euclidean transformations, allowing it to learn invariant atom and edge representations while preserving the equivariance of atom coordinates. This transformer can be used to learn molecular representations robust to geometric transformations. We evaluate the performance of our model through experiments and comparisons with existing methods, showing its ability to generate more stable and valid molecules with good properties. Our model is a promising approach for designing molecules with desired properties and can be applied to a wide range of tasks in molecular modeling.