Soil is the largest terrestrial pool of carbon (C), storing 1395-2293 Pg C, but climate change may cause a large amount of the stored C to be released into the atmosphere as CO2. Accurately predicting how much soil C will be lost or gained in the coming decades to centuries requires improved model structure and parameterization. The current generation of Earth System Models (ESMs) have shown a poor ability at predicting the current spatial patterns of soil C storage. It has recently been found that replacing the conventional soil C model with one that explicitly represents soil microbes leads to major improvements in accuracy in predicting spatial patterns of soil C storage. Additionally, it is well known that improvements in the predictive accuracy of models can be made by calibration of model parameters against datasets. In this study, we first evaluated two soil microbial models with two pools as in German et al. 2012 (GCB, page 1471) and four pools as in Allison et al. (2010, Nature Geosci., Supplementary Methods, pages 1-3). With original parameter values, the two models predicted global soil C between 32 Pg and 6590 Pg C. Then we calibrated the two models against the global observed soil C data using Bayesian data assimilation technique for parameter estimation. The constrained 2-pool model estimated global soil C to be 2100-2500 Pg C with the maximum likelihood value at 2275 Pg C; and the 4-pool model had 2050-2350 Pg C with the maximum likelihood value at 2175 Pg C, which was within the range reported in the literature for the observed global soil C content. Spatial correlation between observed and predicted soil C also improved: before calibration, the 2-pool model explained 29% of variability and the 4-pool model explained 34% of the variability in observed soil C; and after parameter estimation the 2-pool model explained 60% and the 4-pool model explained 51% of variability. Finally, after parameter estimation the uncertainty of cumulative soil C change over 95 years under RCP8.5 scenario reduced by 81% for the 2-pool model and 87.5% for the 4-pool model. For both model formulations RCP8.5 scenario resulted in cumulative soil C loss from 300 to 375 Pg C between the years 2000 and 2100, or 13.5-16% from the estimated global total in 2100. Thus, our results indicated that (i) Bayesian parameter estimation using observed global soil C substantially reduced the uncertainty in parameters, and, hence, in projected soil C dynamics produced by the two models; and that (ii) once calibrated, both microbial models projected large soil C losses.