Abstract Learning to hash has become a crucial technique for big data analytics. Among existing methods, supervised learning approaches play an important role as they can produce compact codes and enable semantic search. However, the size of an instancepairwise similarity matrix used in most supervised hashing methods is quadratic to the size of labeled training data, which is very expensive in terms of space, especially for a large-scale learning problem. This limitation hinders the full utilization of labeled instances for learning a more precise hashing model. To overcome this limitation, we propose a class-wise supervised hashing method that trains a model based on a class-pairwise similarity matrix, whose size is much smaller than the instance-pairwise similarity matrix in many applications. In addition, besides a set of hash functions, our proposed method learns a set of classwise code-prototypes with active bits for different classes. These class-wise code-prototypes can help to learn more precise compact codes for semantic information retrieval. Experimental results verify the superior effectiveness of our proposed method over other baseline hashing methods.