In recent years, with the rapid development of neural networks, visual object tracking is becoming increasingly important in real-world applications such as camera drones and driving assistant systems, even self-driving technology. Neural network-based Siamese trackers achieve satisfactory accuracy in the object tracking field and stand out. However, high-performance Siamese trackers are often designed to be complex and heavy, which hinders their application on resource-limited mobile devices. Thus, compressing the neural network-based tracker to make it lightweight and efficient without obvious performance degradation is of great significance. Inspired by the outstanding work of RepVGG which focuses on compressing multi-branches neural networks without any accuracy cost, we propose the RepSiamses Tracker (RST) which is extremely lightweight and achieves very high tracking speed. In RST, the RepVGG-based backbone depth can be adapted to the different target hardware which benefits from the high flexibility of RepVGG. The experimental results on the high-quality benchmark VOT2018 and LaSOT show that RST achieves satisfactory accuracy and extremely high tracking speed on GPU. More impressively, RST is able to run on the CPU at a hyper-real-time of 69 fps and with very little memory cost of 16.4 MB. Such high tracking speed and low memory cost can bridge the gap between academic algorithms and real-world applications.