With the maturation of LiDAR products and the continuous development of intelligent point cloud data processing methods based on deep learning, the environmental perception capabilities of LiDAR systems have been significantly enhanced. However, the modality provided by a single sensor often has limitations. Additionally, the rapidly increasing demands of existing deep learning algorithms in terms of memory consumption and computational complexity have created an insurmountable gap with the limited resource supply of edge computing devices. We propose a lightweight data fusion framework, TEFusion, for multimodal LiDAR-Camera data. Experimental results show that when applied to 3D object detection, TEFusion achieves the lowest memory consumption and inference latency while maintaining competitive detection accuracy compared to other existing 3D detection fusion algorithms.