Model performance is discussed in their full paper, here:

http://cs.stanford.edu/people/karpathy/deepimagesent/devisag...