Traditional Chinese Medicine (TCM) poses distinct challenges for large language models (LLMs) due to its unique theoretical framework and intricate diagnostic processes. We present TCM-R1, an advanced LLM tailored for TCM applications through a three-stage training strategy. First, continuous pre-training on a comprehensive TCM corpus embeds specialized knowledge, encompassing concepts like Yin-Yang and syndrome differentiation. Second, chain-of-thought-guided supervised fine-tuning enhances reasoning for tasks such as syndrome differentiation, disease diagnosis, and medical dialogues, improving transparency and precision. Third, group relative policy optimization boosts diagnostic accuracy and response quality, particularly for complex cases with similar syndromes. Subjective and objective evaluations across multiple TCM tasks, including TCM-SD, TCM-DD, and TCM-Dialog, demonstrate TCM-R1’s effectiveness, with strong performance in metrics like accuracy, Macro-F1, ROUGE-L, and practitioner-assessed professionalism, fluency, and safety. By combining domain-specific expertise with advanced reasoning and optimization, TCM-R1 provides a robust tool for TCM research and practice. However, limitations include potential inaccuracies in reasoning chains and increased response times, which may affect real-time applications.