Abstract: Large Language Model (LLM) has demonstrated remarkable capabilities across a wide range of domains, inspiring the development of multimodal molecular language models that can jointly process molecular structures, chemical representations, and natural language descriptions. Despite recent progress, existing 3D graph-based molecular models still largely rely on purely data-driven molecular encoders. These models often neglect explicit chemical priors, such as element-level relationships, periodic properties, and chemically meaningful semantic associations, which limits their ability to generalize across the vast chemical space. As a result, they may struggle to accurately capture complex structure–property relationships and to support robust molecule-text reasoning. To address these limitations, we propose KG-MoLM3D, a Knowledge Graph (KG)-Enhanced 3D Molecular Language Model for comprehensive molecular understanding. KG-MoLM3D seamlessly integrates structural, relational, and textual information within a unified multimodal framework. Specifically, we design a KG-enhanced EGNN encoder to fuse 3D molecular geometries with element-level chemical knowledge derived from curated knowledge graphs. In addition, the BLIP-2 framework is employed to align molecular graph representations with textual descriptions, enabling effective cross-modal interaction between molecules and language. Furthermore, we introduce MixLoRA, an efficient parameter adaptation strategy for multitask learning. KG-MoLM3D achieves consistent advantages across molecular representa-tion, molecule-text retrieval, molecule captioning, and property QA tasks; for example, it obtains the lowest MAE on all six MoleculeNet regression datasets, and achieves 96.85% recall@20 in PubChem M2T retrieval and 53.89 METEOR in descriptive property QA.