Clothing keypoint detection is a fundamental task in computer vision with widespread applications in large-scale e-commerce. Existing methods often rely on increasing model size to improve accuracy, making it difficult to achieve a favorable balance between precision and efficiency. This paper introduces a novel framework based on Quantized self-Attention and learnable Coordinate Classification (QACC), which integrates a lightweight convolutional backbone with a quantized self-attention mechanism. By discretizing feature representations into a codebook of visual tokens, QACC-Net effectively captures global dependencies while reducing computational complexity, thereby enabling robust feature modeling despite clothing deformations and occlusions. To further improve localization, a coordinate classification strategy with learnable spatial distributions is introduced, allowing adaptive adjustment of prediction granularity in response to the non-uniform distribution of clothing keypoints. Experiments on the DeepFashion2 dataset demonstrate that QACC-Net achieves competitive accuracy at only 1.9 GFLOPs with 473 FPS inference speed, establishing an optimal trade-off between accuracy and efficiency.