The reputation model is one of the most important methods that can be used to construct trust between peers in peer-to-peer systems. However, almost all reputation models for P2P applications are purely decentralized.They have many defects such as slow convergence speed of trust in node, complicated trust management and overwhelming network cost. So to solve these problems TLT (two-level trust), a two-level trust model, is proposed in this paper. In TLT a series of trust clusters are spontaneously formed that are the minimum unit of trust evaluation. Every trust cluster includes some members and a cluster header. There is a mutual trust relationship between the cluster header and member node. For example, in order to increase inter-cluster service trust the cluster header checks the service performance of members and eliminates malicious members by using the concept of intra-cluster service trust; while member nodes, aiming to heighten the service reputation and receive good quality ervices, also examine the management capability of the cluster header and isolate the malicious cluster headers by employing the concept of proxy trust. Analyses and simulations show that malicious behaviors can be quickly identified in TLT because of the fast convergence speed of trust value and TLT is scalable because of its simple trust management and small network overhead.