XFeat is a lightweight and efficient architecture for resource-efficient visual correspondence, designed to meet the critical need for fast and robust algorithms suitable for resource-limited devices. The method revisits fundamental design choices in convolutional neural networks (CNNs) to detect, extract, and match local features. XFeat is designed to be hardware-agnostic, offering the choice of matching at the sparse or semi-dense levels, each suitable for different downstream applications such as visual navigation and augmented reality. It is the first to efficiently perform semi-dense matching, leveraging a novel match refinement module that relies on coarse local descriptors. XFeat is versatile and hardware-independent, surpassing current deep learning-based local features in speed (up to 5x faster) with comparable or better accuracy, as demonstrated in pose estimation and visual localization. The method is versatile and hardware-independent, and it can run in real-time on an inexpensive laptop CPU without specialized hardware optimizations. The code and weights are available at www.verlab.dcc.ufmg.br/descriptors/xfeat_cvpr24.XFeat is a lightweight and efficient architecture for resource-efficient visual correspondence, designed to meet the critical need for fast and robust algorithms suitable for resource-limited devices. The method revisits fundamental design choices in convolutional neural networks (CNNs) to detect, extract, and match local features. XFeat is designed to be hardware-agnostic, offering the choice of matching at the sparse or semi-dense levels, each suitable for different downstream applications such as visual navigation and augmented reality. It is the first to efficiently perform semi-dense matching, leveraging a novel match refinement module that relies on coarse local descriptors. XFeat is versatile and hardware-independent, surpassing current deep learning-based local features in speed (up to 5x faster) with comparable or better accuracy, as demonstrated in pose estimation and visual localization. The method is versatile and hardware-independent, and it can run in real-time on an inexpensive laptop CPU without specialized hardware optimizations. The code and weights are available at www.verlab.dcc.ufmg.br/descriptors/xfeat_cvpr24.