To solve the problem of relevance model based on unstable clustering estination and its effect on retrieval performance, a new Stable Semantic Relevance Model (SSRM) was proposed. The feedback data set was first formed by using the top N documents from user initial query, after the stable number of semantic clusters had been detected, SSRM was estimated by those stable semantic clusters selected according to higher user-query similarity. Finally, the SSRM retrieval performance was verified by experiments. Compared with Relevance Model (RM), Semantic Relevance Model (SRM) and the clustering-based retrieval methods including Cluster-Based Document Model (CBDM), LDA-Based Document Model (LBDM) and Resampling, SSRM has improvement of MAP by at least 32.11%, 0.41%, 23.64%,19.59%, 8.03% respectively. The experimental results show that retrieval performance can benefit from SSRM.