In order to meet the requirement of high spectrum efficiency of MIMO systems, the number of antennas has been greatly increased to support simultaneous communication of multiple users, and higher order modulation is expected to be employed in the system at the same time, which make signal detection a difficult task. However, most traditional algorithms reply on the channel state information (CSI) accuracy get by channel estimation at receivers. DeepSIC, developed in the existing literature, is a promising framework that introduces data-driven deep learning methods into iterative soft interference cancellation (SIC) algorithms that can address challenges of unknown CSIs and improved performance. In this paper, a signal detection algorithm is designed for MIMO systems under high-order modulation when CSI is unknown, called ResLight-DeepSIC. ResLight-DeepSIC is based on DeepSIC, introducing the idea of residual network (ResNet), and achieves betterdetection performance and lower complexity with lightweight network. Link-level simulation shows that the improved algorithm has a gain of about 1 dB on the symbol error rate (SER), and reduces complexity by about 37% and 63% under 16 QAM and 64 QAM modulation, respectively. Therefore, the ResLight-DeepSIC proposed in this article has advantages in both performance and complexity with large multiple stream transmission and high order modulation.