Early human-designed CNN architectures followed simple design heuristics, leading to models that are powerful but heavy in computation cost and parameter size. With efficiency constraints in mind, later research has focused on compact CNN architectures that are equally powerful but require far less computation. In this chapter, we will discuss three families of well-known compact model architectures, namely SqueezeNet, MobileNet, and ShuffleNet. We will illustrate their design principles and discuss how these designs enable the efficiency improvements in these models.