Abstract: In response to the centralized procurement policy for medical consumables, to ensure that pharmaceutical companies can minimize costs while maintaining hospital demands, this paper designs a medical alliance material distribution optimization algorithm in collaboration with pharmaceutical companies and conducts a pilot project. The algorithm was based on MTL model. By constructing a fair federated training strategy and introducing cosine similarity, the fairness of model training was balanced, thereby maintaining the independence of data from each downstream hospital. To reduce the operation time of the algorithm, the function dimension was reduced by defining the gradient inner product matrix. Based on the drug and unloading preferences of each hospital, a preference term was introduced to form a linear preference strategy, ensuring the uniqueness of the distribution plan. The method proposed in this paper was tested on a self-owned dataset of 36700 entries. The test results show that the accuracy, fairness, and convergence time of the model are all better than similar models, with the model accuracy maintained between 0.898 and 0.935. Practical application results show that the model can provide reasonable assembly plans in multiple scenarios, meet the drug distribution needs of different hospitals, and improve the distribution compatibility between hospitals and enterprises.