Cow keypoint detection is important in research fields such as cow body measurement, behavior recognition, and weight estimation. However, existing deep learning methods for cow keypoint detection still suffer problems such as high network complexity and slow detection speed. A lightweight cow keypoint detection model SimCC-ShuffleNetV2 was proposed. In this model, ShuffleNetV2 was used as the backbone for feature extraction to achieve network lightweight. SimCC was used as the head to achieve keypoint position prediction. SimCC adopted a coordinate classification method that was simple and efficient. To validate the effectiveness of the model, cow keypoints and skeleton structures were designed, and 3600 images were annotated for training and testing. Experimental results showed that the SimCC-ShuffleNetV2 model got an AP50:95 of 88.07%, FLOPs of 1.5×108, parameters of 1.31×106, and detection speed of 10.87f/s, achieving accurate and efficient detection of cow keypoints. Experimental comparisons with the regression-based DeepPose and Heatmap-based HRNet networks demonstrated that SimCC-ShuffleNetV2 got a good balance between accuracy and speed. Moreover, different backbones and detection heads were replaced to verify the influence of different modules on model performance. And the proposed model achieved the best results in all experiments, demonstrating that the combination of ShuffleNetV2 and SimCC had good keypoint detection performance. The model was applied to extract skeleton sequences from four different action videos of cows, and the ST-GCN network was used to classify the four videos, achieving an 84.56% classification accuracy, which indicated that the proposed SimCC-ShuffleNetV2 model was a good keypoint extractor and could provide key information support for tasks such as cow action recognition.