The context word vectors are averaged and passed through a softmax layer. CBOW is faster to train and works well for frequent words.