Pretrained language models have been used in various natural language processing applications. In the mental health domain, domain-specific language models are pretrained and released, which facilitates the early detection of mental health conditions. Social posts, e.g., on Reddit, are usually long documents. However, there are no domain-specific pretrained models for long-sequence modeling in the mental health domain. This paper conducts domain-specific continued pretraining to capture the long context for mental health. Specifically, we train and release MentalXLNet and MentalLongformer based on XLNet and Longformer. We evaluate the mental health classification performance and the long-range ability of these two domain-specific pretrained models. Our models are released in HuggingFace.