Accurately representing permafrost thaw processes in Earth System Models (ESMs) is critical for predicting future climate-carbon feedbacks. While gradual top-down thaw schemes are standard in most of the CMIP6 models, large-scale land surface models increasingly incorporate excess ground ice physics to capture abrupt thaw and subsequent land subsidence. These parameterizations, however, are typically developed and evaluated within single-model frameworks making it difficult to understand how underlying model structure influences the model performance. This study integrates an identical excess-ice ground ice physics into the Common Land Model (CoLM, version 2014) and evaluates it against the Community Land Model (CLM, version 5.1). Using paired historical sensitivity experiments (1901−2014) with and without excess ice across the pan-Arctic, we investigate the effects of excess ice on physical and biogeochemical properties. In this study, we evaluate three configurations: CLM, standard CoLM, and CoLM coupled with a dynamic global vegetation model (CoLM-DV). Our results show that while all configurations capture key physical processes of excess ice such as subsurface latent heat buffering and progressive ice loss under warming, the magnitude and spatial distribution of these responses diverge substantially. The CoLM simulates a much higher sensitivity to warming than CLM, undergoing widespread excess ice loss than the CLM. Biogeochemical responses diverge even more i.e. excess ice melt suppresses regional GPP in CLM (by −40 to −70 TgC yr^−1) and standard CoLM (by −5 to −20 TgC yr^−1) due to warming anomalies maintained by latent heat absorption. On the other hand, CoLM-DV simulates a positive GPP response (+15 to +25 TgC yr^−1) driven by localized warming and wettening that stimulates shrub and grass expansion over bare soil. This inter-model divergence demonstrates that identical physical parameterizations create significantly different climate-carbon feedback trajectories depending on host-model structure and vegetation feedbacks. This work highlights cross-model benchmarking as a vital tool for robust parameterization development across the permafrost modeling community.