In this paper we consider a decentralized maintenance policy for a multi-state system (MSS) consisting of heterogeneous components under a POMDP framework. Components in MSS are classified into low priority components (LPCs) and high priority components (HPCs). Different components may have different deteriorated performances at the same time, and hence the behavior of the system needs to be modeled as a MSS. Two maintenance teams are involved in the reliability predications and maintenance actions for MSS. One team takes the Markov method to predict the deterioration trend of components, and carries out the maintenance action with a lower cost but longer time. The other team estimates the status of MSS based on the sample data which is stochastically related to the system condition. This team can perform the maintenance tasks efficiently within a shorter time but at a higher price. Either maintenance team can be chosen based on the status of system within some additional restraints, such as contract period. Illustrative numerical comparisons are provided at last to show significant cost savings of the decentralized maintenance model over a centralized one.