We present a class of extremely efficient CNN models, MobileFaceNets, whichuse less than 1 million parameters and are specifically tailored forhigh-accuracy real-time face verification on mobile and embedded devices. Wefirst make a simple analysis on the weakness of common mobile networks for faceverification. The weakness has been well overcome by our specifically designedMobileFaceNets. Under the same experimental conditions, our MobileFaceNetsachieve significantly superior accuracy as well as more than 2 times actualspeedup over MobileNetV2. After trained by ArcFace loss on the refinedMS-Celeb-1M, our single MobileFaceNet of 4.0MB size achieves 99.55% accuracy onLFW and 92.59% TAR@FAR1e-6 on MegaFace, which is even comparable tostate-of-the-art big CNN models of hundreds MB size. The fastest one ofMobileFaceNets has an actual inference time of 18 milliseconds on a mobilephone. For face verification, MobileFaceNets achieve significantly improvedefficiency over previous state-of-the-art mobile CNNs.