Text clustering belongs to the unsupervised machine learning, the discriminability of class attributes cannot be measured in clustering. And the traditional text feature selection methods cannot effectively solve the high-dimensional problem. To overcome the weakness in existing feature selection, this paper proposes a new method which introduces the cloud model theory into feature selection, constructs the clouds filter for clustering documents. The distribution of document words is constructed in a microcosmic level. By employing the cloud model digital characteristics we can better compute the separability between feature words. Experimental results with K-means algorithm show that our method can remarkably improve the accuracy of text clustering.