This paper proposes a word-class-based Chinese language model for Mandarin speech recognition with very large vocabulary. The word classes used are developed based on the special structure of Chinese words. We have also developed some improved techniques. The ambiguous syllable filter can delete many confusion syllables and increase significantly the accuracy. The short-term cache memory can help the language model to adapt to the current application domain, and the learning module can significantly reduce the zero values in the language model.