Automatically and accurately recommending discharge medications for patients based on unstructured Chinese electronic medical records (EMRs) remains a major challenge for current clinical decision support systems. An effective recommendation model can help standardize treatment practices and reduce discrepancies in therapeutic strategies caused by variations in physicians’ interpretations. For the Evaluation Task of the 11th China Health Information Processing Conference (CHIP 2025), this paper proposes a medication recommendation method built upon the Qwen2.5-7B-Instruct model, incorporating multi-format fine-tuning and an optimized ensemble voting strategy. Specifically, we design two output formats, a structured list format and a natural language format, and apply LoRA for efficient parameter adaptation. Based on these settings, three candidate predictions are generated using different decoding temperatures and integrated through an optimized voting mechanism that enhances recall, where medications predicted by at least two models are retained, and in cases of complete disagreement, the union of all predicted medications is preserved. Experimental results show that our method achieves an accuracy of 53.78%, ranking fourth in the competition, thereby demonstrating its effectiveness and robustness.