Purpose: Establishing a multi-level technical topic branching system serves as foundational work to address information explosion and the value of knowledge. This study proposes a novel framework integrating generative AI and density-based clustering to address hierarchical topic discovery in patent analytics, overcoming limitations of static taxonomies in existing methods. Methodology: We propose an automated hierarchical topic learning framework for large-model patents, comprising four core modules:(1) Domain-Specific Data Curation: A retrieval formula integrating IPC codes and domain keywords was designed to collect 43,355 raw patents from incoPat. After deduplication and text fusion into a unified “patent description” field), the LargeModelPatent-2024 dataset (34,985 patents) was constructed. (2) Recursive Density Clustering: (i) Vectorization & Dimensionality Reduction: Sentence-BERT generated 512-D vectors, compressed to 2-D via UMAP (n_neighbors = 15, cosine metric); (ii) Adaptive Clustering: HDBSCAN (min_cluster_size = 100) performed primary clustering, yielding 14 high-density clusters (silhouette = 0.366). For data-intensive clusters, recursive secondary clustering generated 21 leaf-node topics;(3) Generative Topic Labeling: TF-IDF extracted keywords per cluster, weighted and input to Wenxin API to generate interpretable labels; (4) Classification & Deployment: A BERT classifier (95.02% accuracy) trained on leaf-node topics powered a Django-Vue web system for real-time patent analysis. Results: The approach achieves a silhouette coefficient of 0.366, outperforming IPC classification in granularity. The auto-generated topic labels show 89% consistency with expert evaluation. The operational system supports batch processing and real-time topic navigation. Conclusion: This fusion of generative AI and density clustering enables dynamic, hierarchical patent topic mining, providing a scalable solution for technology surveillance. The LargeModelPatent-2024 dataset and code are open-sourced for reproducibility.