Knowledge graphs have become essential tools for structuring and representing complex information across various domains, enabling advanced applications such as semantic search, recommendation systems, and automated reasoning. The novel concept introduced in this paper involves enhancing multi-modal knowledge graph completion through fine-grained tokenization of discrete modality information, significantly improving the richness and utility of knowledge graphs. We modified the Mistral Large model to integrate and process diverse data modalities, including text, images, and audio, via separate encoders and a fusion layer for integrated representation. The modified model demonstrated superior performance in multi-modal knowledge graph completion tasks, as evidenced by improved precision, recall, and F1-scores compared to baseline models. The rigorous training regimen, which combined supervised and unsupervised learning techniques, optimized the model's ability to generate accurate and contextually rich knowledge graphs. Comprehensive evaluations highlighted the robustness, scalability, and generalization capabilities of the enhanced model, underscoring its potential for broad applicability in domains requiring sophisticated knowledge representation and reasoning capabilities.