Named entity disambiguation (NED) is an important stage in Natural Language Processing (NLP) which automatically resolves mentions to entities in a given knowledge base (KB) like Wikipedia. NED is a complex and challenging problem due to the inherent ambiguity between real world mentions and the entities they refer to. Most existing studies use hand-crafted features to represent mentions, context and entities, which is labor intensive. In this paper, we address this problem by presenting a new NED model which combining local, context and global evidence. By leveraging the learned mixed dense word-level and topic-level representations and the graph-based disambiguation approach, context and global features are well captured. Experiments for NED are conducted on AIDA dataset, which show that the proposed model can obtain state-of-the-art results.