Mid-infrared reflectance spectroscopy (MIRS, 4000–400 cm−1) is being considered to provide accurate estimations of soil properties, including soil organic carbon (SOC) and soil inorganic carbon (SIC) contents. This approach has mainly been demonstrated by using datasets originating from the same area A, with similar geopedological conditions, to build, validate and test prediction models. The objective of this study was to analyse how MIRS performs when applied to predict SOC and SIC contents, from a calibration database collected over a region A, to predict over a region B, where A and B have no common area and different soil and climate conditions. This study used a French MIRS soil dataset including 2178 topsoil samples to calibrate SIC and SOC prediction models with partial least squares regression (PLSR), and a Tunisian MIRS topsoil dataset including 96 soil samples to test them. Our results showed that when using the French MIRS soil database, i) the SOC and SIC of French validation samples were successfully predicted using global models (R2 val = 0.88 and 0.98, respectively), ii) the SIC of Tunisian samples was also predicted successfully both using a global model and using a selection of spectral neighbours from the French calibration database (R2 test of 0.96 for both), iii) the SOC of Tunisian samples was predicted moderately well by global model (R2 test of 0.64) and a transformation by natural logarithm of the calibration SOC values significantly improved the SOC prediction of Tunisian samples (R2 test of 0.97), and iv) a transformation by natural logarithm of SOC values provided more benefit than a selection of spectral neighbours from the French calibration database for predicting Tunisian SOC values. Therefore, in the future, MIRS might replace conventional physico-chemical analysis techniques, or at least be considered as an alternative technique, especially when optimally exhaustive calibration databases will become available.