Semantic matching is a fundamental task in the field of natural language processing (NLP). Existing single model can’t deal with the problems of sparse features, semantic ambiguity and lack of semantics in colloquial data well in social media. Therefore, we design a hybrid model based on multi-level external knowledge (HME) for Chinese semantic matching. Firstly, HME introduces Cilin as an external knowledge source, and performs feature expansion for text pairs with high literal similarity according to different parts of speech. Secondly, HME introduces the HowNet as an external knowledge source to solve the problem of semantic ambiguity. Thirdly, HME introduces clickthrough data as an external knowledge source, and enhances the context representation for a small number of text pairs with unclear matching results to improve the accuracy of model matching. The experimental results of the HME on two large public Chinese datasets, LCQMC and BQ, show that HME has better performance than most existing models.