This paper proposes a discrete mathematics teaching platform based on a multimodal learning resource repository and the DeepSeek-RAG integration engine. The multimodal repository collects diverse data related to discrete mathematics, which undergoes preprocessing, annotation, and deep learning-based optimization (CNNs for images, RNNs for text or speech). Knowledge is stored using a dual-structure approach combining knowledge graphs and vector databases, enabling semantic association, intelligent retrieval, and inference of multimodal information. The DeepSeek-RAG integration engine combines retrieval-augmented generation (RAG) with the DeepSeek large language model to perform efficient sparse or dense retrieval, accurately locating relevant content within the multimodal repository. Through multimodal fusion, redundancy reduction, and enhanced relevance modeling, the engine generates results with improved contextual consistency, knowledge richness, and accuracy, effectively supporting retrieval, question answering, and assessment functionalities. Based on this platform, we propose a hybrid teaching model that integrates traditional instruction, intelligent platforms, and collaborative group learning, which has demonstrated significant effectiveness in practical teaching scenarios.