Reflecting recent advances in our understanding of soil organic carbon (SOC) turnover and persistence, a new generation of models increasingly makes the distinction between the more labile soil particulate organic matter (POM) and the more persistent mineral-associated organic matter (MAOM). Unlike the typically poorly defined conceptual pools of traditional SOC models, the POM and MAOM pools can be directly measured for their carbon content and isotopic composition, allowing for pool-specific data assimilation. However, the new-generation models’ predictions of POM and MAOM dynamics have not yet been validated with pool-specific carbon and 14C observations. In this study, we evaluate 5 influential and actively developed new-generation models (CORPSE, Millennial, MEND, MIMICS, SOMic) with pool-specific and bulk soil 14C measurements of 77 mineral topsoil profiles in the International Soil Radiocarbon Database (ISRaD). We find that all 5 models consistently overestimate the 14C content (Δ14C) of POM by 67‰ on average, and 3 out of the 5 models also strongly overestimate the Δ14C of MAOM by 74‰ on average, indicating that the models generally overestimate the turnover rates of SOC and do not adequately represent the long-term stabilization of carbon in soils. These results call for more widespread usage of pool-specific carbon and 14C measurements for parameter calibration, and may even suggest that some new-generation models might need to restructure their simulated pools (e.g. by adding inert pools to POM and MAOM) in order to accurately reproduce SOC dynamics.