Abstract: Vertical mixing parameterization is a major source of uncertainty in ocean circulation models. Recently, deep learning has been used to improve traditional parameterization schemes as it can effectively capture complex nonlinear processes. This study employs hybrid programming technology to integrate the DLVMP (deep learning-based vertical mixing parameterization) proposed by Fang et al. (2025) into the LASG/IAP Climate System Ocean Model (LICOM) developed by the Institute of Atmospheric Physics. Three long-term climate simulations were conducted: A control experiment CNTR, using Canuto et al. (2001) scheme, a K-Profile Parameterization (KPP) experiment, and a DLVMP sensitivity experiment. The results show that DLVMP shares some biases with KPP but improves the simulation of equatorial subsurface temperatures by using observations from that region. Compared with the Canuto et al. (2001) scheme, both DLVMP and KPP underestimated surface mixing and overestimated subsurface mixing in the equatorial Pacific, resulting in significant differences in temperature, salinity, mixed-layer depth, Atlantic Meridional Overturning Circulation (AMOC), and the STC (subtropical cell) among the experiments. Additionally, the underestimated surface vertical mixing weakens the mid-latitude ventilation process, resulting in an upward shift and intensification of the STC in the DLVMP experiment, accompanied by a significant cold temperature bias near the thermocline. This study highlights the challenges of considering the appropriate size and quality of the training dataset for deep learning-based parameterizations, providing key insights for future improvements.