In this paper, a length distribution model for intonational phrase prediction is proposed. This model presents the probabilities that a certain length sentence is divided into some certain length intonational phrases. We will discuss how to estimate the probabilities in the model from training corpus, and how to apply it to intonational phrase prediction. We combine this model with a maximum entropy model which implements local context information. Experiment results show that length distribution is valuable information for intonational phrase prediction, and that it is able to make significant extra contribution over the maximum entropy model in terms of average score and unacceptable rate.