Conventional Q10 soil organic matter decomposition models and more complex microbial models are available for making projections of future soil carbon dynamics. However, it is unclear (1) how well the conceptually different approaches can simulate observed decomposition and (2) to what extent the trajectories of long-term simulations differ when using the different approaches. In this study, we compared three structurally different soil carbon (C) decomposition models (one Q10 and two microbial models of different complexity), each with a one- and two-horizon version. The models were calibrated and validated using 4 years of measurements of heterotrophic soil CO2 efflux from trenched plots in a Dahurian larch (Larix gmelinii Rupr.) plantation. All models reproduced the observed heterotrophic component of soil CO2 efflux, but the trajectories of soil carbon dynamics differed substantially in 100 year simulations with and without warming and increased litterfall input, with microbial models that produced better agreement with observed changes in soil organic C in long-term warming experiments. Our results also suggest that both constant and varying carbon use efficiency are plausible when modeling future decomposition dynamics and that the use of a short-term (e.g., a few years) period of measurement is insufficient to adequately constrain model parameters that represent long-term responses of microbial thermal adaption. These results highlight the need to reframe the representation of decomposition models and to constrain parameters with long-term observations and multiple data streams. We urge caution in interpreting future soil carbon responses derived from existing decomposition models because both conceptual and parameter uncertainties are substantial.