The objective of this work was to evaluate the robustness of Near Infrared (NIR) and Mid-Infrared (MIR) models on a set of apricots with large phenotypic variability (eight contrasted cultivars, four maturity stages, and 180 hybrids from a cross between 'Goldrich' and 'Moniqui'). The error of prediction obtained by NIR models on intact fruit was 9.2% for soluble solids content (SSC) and 16.5% for titratable acidity (TA). Using MIR models on slurries, this error was 6.1% for SSC and 8.6% for TA. Overall, the prediction performance of models established using large phenotypic variability was stable enough for global quality traits such as SSC and TA.