In this paper, we explore the effectiveness of deep learning models for text sentiment classification in Arabic. We propose the evaluation of Deep Belief Networks and deep Auto Encoders models. Three architectures are derived using the selected Arabic data set. The deep learning models are trained with features based on the recently developed Arabic Sentiment Lexicon (ArSenL) in combination with other standard lexicon features. The models are tested and compared to support vector machines (SVM) models. The evaluation is carried out using Linguistic Data Consortium Arabic Tree Bank dataset. The results show potentials for the chosen deep learning models, but need further improvements to provide performances superior to of conventional SVM. The paper sets directions for exploration of other and improved deep learning neural network architectures.