Automated tooth instance segmentation on dental radiographs is a crucial step in establishing digital dental workflows. However, unlike the realm of natural images, there is currently no visual foundation model that can implement tooth instance segmentation accurately. In this paper, we built the first visual foundation model, SemiT-SAM, for tooth instance segmentation. This foundation model was meticulously designed in terms of model architecture design, the training data corpus, and the semi-supervised learning strategy. The SemiT-SAM inherited the capability of the SAM and was trained on a large-scale dataset TSI15k via the label-guided teacher-student knowledge distillation strategy. Based on SemiT-SAM, we participated in the challenge of MICCAI STS 2024: Panoramic X-ray Images, and achieved satisfying performance with scores of 90.52% (image-level NSD) and 86.89% (image-level Dice) on the validation set. The checkpoint and code of SemiT-SAM, as well as the training dataset TSI15k, are available at: https://github.com/isbrycee/SemiTNet .