In typical partial label learning, each training sample corresponds to a candidate set with only one true label. Eliminating the ambiguous candidate set is the core task of partial label learning. The maximum margin method is a popular paradigm to solve partial label learning problem. It is an effective attempt to directly optimize the margin between the real label and the rest of the labels. However, recent studies have found that the optimal margin distribution is the key to better generalization performance rather than maximizing the minimum margin. In this paper, we propose a novel approach, named PL-ODM (Partial Label Optimal Margin Distribution Machine), which tries to disambiguate the candidate label sets and achieve optimal margin distribution simultaneously. First of all, the mean and variance are used to describe the margin distribution. After that, an alternating optimization process is used to derive the true label information and optimize the margin distribution. By observing the experimental results on the real data set and the controlled UCI data set, the effectiveness of our proposed method is fully demonstrated.