Abstract: An efficient convolutional operator named Quark Module for multi-object tracking (MOT) is proposed for the purpose of lightweight, which can be used to lightweight deep neural network models from both width and depth. Based on Quark Module, a lightweight MOT model QuarkJDE was proposed to improve the classical JDE (Joint Detection and Embedding) algorithm. In order to verify the generalization ability of Quark Module, another lightweight classification model QuarkNet was constructed. Diverse experiments were carried out on basis of public benchmark datasets to test the above two models, all the results fully proved the efficiency and feature learning ability of Quark Module.