Dialogue systems which can communicate with people in natural language is popularly used in entertainments and language learning tools. As the development of deep neural networks, Sequence-to-Sequence models become the main stream models of conversation generation tasks which are the key part of dialogue systems, because Sequence-to-Sequence models is good at dealing with the tasks like machine translation and conversation generation whose input's length and output's length is unknown previously. However, recent works find that Sequence-to-Sequence models tend to respond in dull sentences. We propose a clustering based adaptive Sequence-to-Sequence model to improve the performance of dialogue systems. Different with previous models who treat all the dialogue data as input of a single model, we cluster the dialogue data and use several Sequence-to-Sequence models to train different cluster of data to catch different characteristic in different cluster. Our experiments show that our models can improve the performance of dialogue systems. Copyright © 2018 by The Institute of Electrical and Electronics Engineers, Inc.