The performances of CMIP5 and CMIP6 Earth System models in simulating China’s terrestrial vegetation biomass were assessed based on two sets of the observation biomass carbon density data, the first to ninth (1950-2018) China Forest Resource Inventory data and a set of remote sensing land cover data. The impacts of different Land-Use Harmonization (LUH) datasets on vegetation biomass simulations of CMIP5 and CMIP6 models were further explored. Our results show that both CMIP5 and CMIP6 models overestimate China"s total vegetation carbon storage during 1995-2004 (28.0 ± 6.0 Pg C for CMIP5, 25.3 ± 7.7 Pg C for CMIP6, and the two sets of reference data are 18.1 Pg C and 18.7 Pg C respectively). The spatial distributions of vegetation carbon density as simulated in CMIP6 models are better than those in CMIP5, with improved Taylor Skill Score(TSS) values and reduced model uncertainties. During 1950-1990, China’s vegetation in CMIP5 and CMIP6 are carbon sources of ?89.4 Tg C yr-1 and ?58.2 Tg C yr-1, respectively. These carbon sources significantly increase to ?256.6 Tg C yr-1 (CMIP5) and ?171.0 Tg C yr-1(CMIP6) in the 1980s. During 1990-2014, in CMIP5 the vegetation carbon sources are much weaker (?48.1 Tg C yr-1) , while in CMIP6 those sources are changed to weak sinks (42.8 Tg C yr-1, P<0.05). The differences of China"s vegetation carbon between CMIP5 and CMIP6 are closely related to their different land change conditions. Compared to Land-Use Harmonization (LUH1) in CMIP5, the forest covers and their changes in LUH2(in CMIP6) are more consistent with China Forest Inventory data and are closer to that since 1980s. This study implies that LUH datasets used in CMIPs have large bias from China’s forest and crop trajectories over the past 65 years. More accurate land cover datasets are essential for improving the simulation of vegetation carbon in the future CMIP models.