Abstract: Medical artificial intelligence (AI) has gradually transitioned from algorithmic validation to real clinical deployment. This article focuses on the practical application of AI in medicine. It proposes the necessity of moving from single-model tools to collaborative systems of large and small models. It summarizes the collaborative paradigm into three typical levels: the perception layer, the decision-making layer, and the system layer, thus more closely reflecting real-world clinical workflows. It further analyzes the trustworthiness challenges under the collaborative paradigm, including decision-making conflicts across multiple models, error accumulation and amplification in multi-model collaboration, inconsistencies caused by version and knowledge asynchrony, as well as the prioritization and weighting of heterogeneous evidence sources. Corresponding potential feasible technical paths are also provided.