Abstract: In recent years, large language models have advanced rapidly, with general-purpose models achieving substantial improvements in both scale and capability. However, due to insufficient optimization for medical expertise and the essential competencies required in clinical scenarios, general-purpose large language models often perform poorly when directly applied to the medical domain. Against this backdrop, medical domain-specific large language models fine-tuned from general-purpose foundation models have emerged as a promising solution. This article first discusses the key technical challenges and reviews the related research progress in China and abroad from four perspectives: training data quality control, medical knowledge optimization, scenario-specific capability enhancement, and complex task orchestration. It then presents the technical solutions adopted by Peking University’s “Xiaobei Health” medical domain-specific large language model in these four aspects. Finally, the article introduces the overall performance evaluation and practical applications of the “Xiaobei Health” model. Experimental results on the Chinese medical benchmark (CMB) show that the “Xiaobei Health” model achieves an average accuracy of 87.95%, significantly outperforming existing representative models, demonstrating the effectiveness and practical value of the proposed technical framework.