To address the issues of limited storage and computing resources in existing underwater equipment used in industries such as fishing and seabed exploration, as well as the large size of detection models, which are challenging to operate efficiently on terminal devices, a lightweight underwater biological detection algorithm combining pruning and knowledge distillation is proposed. First, the C2f_GSConv structure is designed to replace the C2f module in the neck of the YOLOv8n network, reducing the overall computational complexity and optimizing the model structure. Second, MPDIoU is introduced to replace CIoU as a new loss function, accelerating the convergence of the regression bounding box and improving detection performance. The LAMP pruning algorithm is then applied to trim the model by removing redundant channel information and convolutional kernels, further reducing the number of parameters and computational complexity, thus compressing the model size. Finally, knowledge distillation is employed to restore the model’s detection accuracy and reduce the precision loss caused by pruning. Experimental results show that, on the URPC dataset, the improved model outperforms the benchmark YOLOv8n model with a 1.8% increase in mAP50, a 62% reduction in parameters, a 56% reduction in computational cost, and a 186 f/s increase. The results also demonstrate excellent performance upon deployment and verification on an embedded development board, confirming its suitability for application in low-configuration underwater equipment.