Mid-Infrared reflectance spectroscopy (MIRS, 4000–400 cm−1) is being considered to provide accurate estimations of soil inorganic carbon (SIC) contents, based on prediction models when the test dataset is well represented by the calibration set, with similar SIC range and distribution and pedological context. This work addresses the case where the test dataset, here originating from France, is poorly represented by the calibration set, here originating from Tunisia, with different SIC distributions and pedological contexts. It aimed to demonstrate the usefulness of 1) classifying test samples according to SIC level based on the height of the carbonate absorbance peak at 2510 cm−1, and then 2) selecting a suitable prediction model according to SIC level. Two regression methods were tested: Linear Regression using the height of the carbonate peak at 2510 cm−1, called Peak-LR model; and Partial Least Squares Regression using the entire MIR spectrum, called Full-PLSR model. First, our results showed that Full-PLSR was 1) more accurate than Peak-LR on the Tunisian validation set (R²val = 0.99 vs. 0.86 and RMSEval = 3.0 vs. 9.7 g kg−1, respectively), but 2) less accurate than Peak-LR when applied on the French dataset (R²test = 0.70 vs. 0.91 and RMSEtest = 13.7 vs. 4.9 g kg−1, respectively). Secondly, on the French dataset, predictions on SIC-poor samples tended to be more accurate using Peak-LR, while predictions on SIC-rich samples tended to be more accurate using Full-PLSR. Thirdly, the height of the carbonate absorbance peak at 2510 cm−1 might be used to discriminate SIC-poor and SIC-rich test samples (<5 vs. > 5 g kg−1): when this height was > 0, Full-PLSR was applied; otherwise Peak-LR was applied. Coupling Peak-LR and Full-PLSR models depending on the carbonate peak yielded the best predictions on the French dataset (R²test = 0.95 and RMSEtest = 3.7 g kg−1). This study underlined the interest of using a carbonate peak to select suitable regression approach for predicting SIC content in a database with different distribution than the calibration database.