HStar is implemented to support large scale OWL documents management. Physical storage model is designed on file system based on semantic model of OWL data. Inference and query are implemented on such physical storage model. Now HStar supports characters of OWL Lite and we try to adopt strategy of partial materializing inference data, which is different from most of existing semantic repository systems. In this paper we first give the data model which HStar supports, then give an analysis of our inference strategy; storage model and query process are discussed in detail; experiments for comparing HStar and related systems are given at last.