Minerals are non-renewable resources that are indispensable for contemporary industrial production. The advent of intelligent ore sorting technologies has been pivotal in enhancing mineral utilization, a phenomenon that has been further propelled by the advent of YOLO (You Only Look Once) object detection models for end-to-end detection. However, intelligent ore sorting equipment deployed at mining sites has more exacting requirements regarding the parameters, computational complexity, and inference speed of deep learning networks. Accordingly, this paper proposes a lighter and faster ore sorting YOLO model, namely LFOS-YOLO, derived from the configuration of the YOLO model. The model incorporates partial convolution, parameter-free attention mechanisms, and redundant channel pruning strategies with the objective of achieving an accuracy-lightweight tradeoff that meets the requirements for deployment in a mining site. The experimental results demonstrate that LFOS-YOLO is capable of accomplishing the task of sorting ore samples from a mine in Liaoning, China, with fewer parameters (1.46M), lower GFLOPs (Giga Floating-Point Operations per Second) (3.4), and higher FPS (Frames per Second) (74), achieving the highest mAP (mean Average Precision) (95.4%), which outperforms other models in the YOLO series.