RepVGG is a simple yet powerful convolutional neural network architecture that combines a VGG-like inference-time body composed of 3 × 3 convolutions and ReLU with a multi-branch training-time topology. The decoupling of the training-time and inference-time architectures is achieved through structural re-parameterization, allowing the model to be converted from one to the other. On ImageNet, RepVGG achieves over 80% top-1 accuracy, outperforming many complex models. It runs 83% faster than ResNet-50 and 101% faster than ResNet-101 while maintaining higher accuracy. RepVGG is designed for efficient inference on GPUs and specialized hardware, offering a favorable speed-accuracy trade-off compared to state-of-the-art models like EfficientNet and RegNet. The code and trained models are available on GitHub.RepVGG is a simple yet powerful convolutional neural network architecture that combines a VGG-like inference-time body composed of 3 × 3 convolutions and ReLU with a multi-branch training-time topology. The decoupling of the training-time and inference-time architectures is achieved through structural re-parameterization, allowing the model to be converted from one to the other. On ImageNet, RepVGG achieves over 80% top-1 accuracy, outperforming many complex models. It runs 83% faster than ResNet-50 and 101% faster than ResNet-101 while maintaining higher accuracy. RepVGG is designed for efficient inference on GPUs and specialized hardware, offering a favorable speed-accuracy trade-off compared to state-of-the-art models like EfficientNet and RegNet. The code and trained models are available on GitHub.