Abstract: To address the challenge that existing CAD model generation methods struggle to achieve both high-quality generation and design intent preservation, particularly showing insufficient capability in handling implicit design constraints, this paper proposes a hierarchical CAD generation method based on discrete diffusion. The method represents CAD models as three-level neural code trees, spanning from local curve geometry to global component arrangement. First, a vector quantized variational autoencoder with adaptive mask skip connections was employed to extract multi-level design features. Then, a hierarchical noise scheduling strategy was combined to dynamically adjust the noise distribution of codebooks at each level. Finally, a two-stage cascaded autoregressive Transformer was utilized to convert code trees into complete CAD command sequences. Extensive experiments on the extended DeepCAD dataset demonstrate that the proposed method exhibits excellent performance in unconditional generation tasks, achieving optimal and sub-optimal results across all metrics. Notably, the method achieves 89.15% in point cloud evaluation metric R_coverage and 0.61 in D_JSD, both surpassing existing state-of-the-art methods. Furthermore, the method shows stronger design intent preservation capability in conditional generation tasks. User studies further validate the effectiveness of the method in maintaining design intent, providing a novel solution for CAD automation.