With the transformation of pig breeding industry to largescale and intensive, non-intrusive individual identification technology is very important for traceback, food safety, disease control and scientific breeding. Pig facial landmark detection serves as a fundamental requirement for achieving non-invasive pig identification. A pig facial landmark detection model named FCM-SimCC was introduced, building upon the SimCC landmark detection algorithm. The model replaced CSPDarkNet with FasterNet for feature extraction and incorporated the CA attention mechanism within FasterNet to enhance the capture of long-distance features. Supervision of the model was achieved through the MLT adaptive weight multi-task loss function combined with KL divergence loss and Wing Loss. Test on a dataset of 4861 images was done, representing a variety of pig breeds and facial poses, the FCM-SimCC model attained mean average precision, 50% average precision, and 75% average precision of 76.12%, 93.44%, and 83.25%, respectively. These results indicated improvements of 3.14, 1.77, and 4.47 percentage points over the original model, with a reduced computational demand to 2.79×109 and a parameter count of 1.38×107, marking a 38.68% decrease in floating-point operations and a 20.16% reduction in parameters. When compared with mainstream landmark detection methodologies such as DeepPose, HRNet, and YOLO X-Pose, the FCM-SimCC model showcased its ability to provide rapid and precise pig facial landmark detection with lower computational resources and fewer parameters, offering valuable insights for similar tasks in pig facial landmark detection and individual pig identification.