Visible, near-infrared and short-wave infrared (VNIR/SWIR, 400-2500 nm) laboratory soil spectrometry is now considered to provide accurate estimations of primary soil properties (clay, calcium carbonate, iron, soil organic carbon, etc.). The performances of primary soil property prediction models are evaluated in regard to figures of merit calculated over calibration and validation databases but not in regard to the spatial extent of predicted soil samples. The objective of this study was to analyze regional model performances for soil property prediction at regional and within-field extents within contrasted representative geopedological situations. This study used a database of 240 soil samples collected over eight vineyard fields located in the Languedoc Region (southern France) (between 20 and 36 soil samples per field) for which VNIR/SWIR laboratory spectra were acquired and two soil physico-chemical properties (clay and calcium carbonate) were measured. Soil property prediction models were built using the classical partial least square regression (PLSR) method, which links the VNIR/SWIR laboratory spectra and the physico-chemical soil property. Our results showed that both clay and calcium carbonate prediction models are accurate at the regional extent, whereas prediction model performances at the within-field extent depend on the model robustness. Therefore, primary soil properties predicted by VNIR/SWIR laboratory spectra must be used with care at different extents.