Besides altering the model architecture with sparsity, an alternative way to improve DNN efficiency is to change the representation of weights and activations in the model, namely quantization. Quantization uses fixed-point representations of a lower precision to represent the weights and activations in a DNN model, so as to reduce the memory consumption and computation cost. In this chapter, we will start with discussing different types of quantization and their impacts, followed by detailed discussion on the training methods to achieve better accuracy and efficiency under linear quantization. We will also cover training-free post-training quantization methods for complicated models. We will discuss multiple research papers as case studies along the way.