Space-Air-Ground Integrated Networks (SAGIN) present stringent and complex requirements for network design and management due to their heterogeneity, self-organizing nature, dynamic characteristics and so on. Despite extensive research on deep learning in the context of SAGIN, challenges such as inflexibility, lack of generalization, limited robustness, and insufficient scalability persist. Large language models (LLMs), with their powerful natural language processing, understanding, and generation capabilities, show significant potential in meeting the complex demands of SAGIN. We thoroughly analyze the application of LLMs in optimizing channel models, improving communication algorithms, and enhancing network management and security performance to address the unique characteristics and requirements of SAGIN. Additionally, it identifies the challenges faced during the deployment of LLMs, including computational resource constraints, data transmission limitations, model updates, real-time processing, and system integration. Corresponding solutions are proposed to ensure the efficient operation and reliability of SAGIN. Through this research, we aim to provide innovative solutions for the intelligent management of SAGIN, promoting the continuous development and application of SAGIN.