Rapid and accurate identification of coal and gangue is essential for improving coal-resource utilization and reducing environmental pollution. Multispectral imaging has been widely used for coal-gangue sorting because of its abundant spectral information. However, unstable illumination and target clutter in workshop environments often limit detection accuracy and robustness. To address these challenges, a lightweight multispectral coal-gangue detection model was proposed for workshop environments, termed learnable dynamic activation-hierarchical context fusion-inner-MPDIoU YOLO (LHIM-YOLO). First, a lightweight learnable dynamic activation module (LDAM) was designed to enhance the modeling of nonlinear differences among spectral bands. Second, a hierarchical context fusion feature pyramid network (HCFFPN) was constructed to efficiently integrate multiscale features and improve the representation of fine-grained textures and macroscopic structures. Finally, an auxiliary-box-constrained inner minimum point distance intersection over union (Inner-MPDIoU) loss was introduced to strengthen geometric constraints, including center-distance and containment relationships, thereby improving localization stability in complex scenes. Experimental results show that LHIM-YOLO achieves an mAP 0.5 of 74.69% with only 3.04×10⁶ parameters. Compared with the benchmark models, it provides superior detection performance while maintaining high accuracy and real-time capability, demonstrating strong potential for practical engineering applications.