Abstract: With the acceleration of digital transformation, large language models (LLMs; a frontier technology in artificial intelligence) are playing an increasingly pivotal role in the intelligent transformation of government services. This study systematically examines the application of language models in the public sector and conducts a comprehensive analysis of their technical architectures, methodological approaches, and practical effectiveness. First, it outlines the fundamental concepts and evolutionary trajectory of LLMs, thereby tracing their development from early language models to the current sophisticated large-scale architectures. It also provides a strong foundation for understanding their governmental applications. Subsequently, the study extensively evaluates the key enabling technologies for LLMs in government contexts. These include model fine-tuning strategies adapted to specific administrative scenarios to enhance task-specific performance, prompt engineering techniques that involve meticulously designed input formulations to guide models toward outputs compliant with governmental standards, knowledge augmentation methods that integrate domain-specific governmental expertise to strengthen model capabilities, and multi-agent collaboration frameworks that enable the coordinated execution of complex administrative tasks. The observations indicate that the “pre-training + domain-specific fine-tuning” paradigm endows government-oriented LLMs with robust generalization capability and adaptability. Well-crafted prompts significantly improve the model's comprehension of official directives, whereas structured knowledge-based constructions reinforce the model’s factual grounding. These technological advancements have driven comprehensive innovation in conventional governance models. This, in turn, has substantially enhanced the efficiency and quality of public services, facilitated cross-departmental data integration and process re-engineering, broken down information silos, and optimized administrative workflows. Furthermore, this study provides a comparative analysis of closed- and open-source LLMs in government applications, thereby evaluating their respective strengths and limitations. Using the DeepSeek large open-source model as a case study, the technical adaptation process and real-world implementation outcomes are detailed. Nevertheless, challenges persist with regard to the deployment of large models in government settings. Data security concerns directly affect the confidentiality of sensitive governmental information and protect citizen privacy. The limited model interpretability undermines the transparency and trustworthiness of the policymaking process. Ethical risks (including algorithmic bias and the potential generation of misleading or false content) pose risks to impartiality and institutional stability, and represent critical obstructions to a broader adoption. Drawing on theoretical insights and empirical evidence, this study proposes future directions for advancing governmental LLM applications, including service optimization; continuous performance enhancement; multimodal data processing capabilities; strengthened collaboration among government, industry, academia, and research institutions; and improvements in regulatory frameworks. These recommendations aim to provide both theoretical guidance and practical references for the intelligent evolution of public administration. Ultimately, these support the realization of efficient, accessible, and intelligent advanced government services and elevate digital governance to a higher level.