3D tooth landmark detection is a critical task in digital orthodontics. Several existing methods rely on predicting a shared heatmap, necessitating post-processing steps like clustering, while others predict a fixed number of heatmaps for corresponding landmarks. These methods struggle to handle a variable number of landmarks, resulting in limited robustness. To address these issues, we introduce a two-stage 3D tooth landmark detection framework that integrates TeethGNN for segmentation and an end-to-end neural network (TL-DETR) for landmark detection. Our TL-DETR is designed with two branches for predicting fixed landmarks and a variable number of cusp landmarks on individual teeth. The fixed landmark branch is supervised by the ground truth heatmaps of the fixed landmarks, while the cusp landmark branch provides a preset number of heatmaps and probabilities for potential cusp points. We employ bipartite matching loss to measure the discrepancy between the prediction sets and the ground truth cusp heatmaps, facilitating end-to-end inference. We perform a comprehensive analysis and evaluation on the 3DTeethLand dataset. The experimental results demonstrate that our method significantly outperforms existing baseline approaches, exhibiting both efficiency and robustness. (The code will be released at https://github.com/bibi547/TL-DETR ).