Milk filtration procedures are gaining relevance in the dairy industry because milk ultra- and nanofiltrates are used to increase milk processing efficiency, and as additives for products with improved nutraceutical properties. This study aimed to develop Fourier-transformed mid-infrared spectroscopy calibrations for ultra- and nanopermeate and retentate fractions of defatted and delactosated milk. A total of 154 samples from different milk fractions were collected and analyzed using reference methods to determine protein, solids-not-fat, glucose, and galactose content. The obtained values were matched with their respective Fourier-transformed mid-infrared spectroscopy spectra to develop new prediction models. Calibrations for each trait were built following 3 different approaches to get the best prediction models: (1) using the entire data set, (2) using 3 subsets based on component concentrations (level approach), and (3) using hierarchical clusters calculated with pairwise Mahalanobis distance among spectra (cluster approach). Calibrations were developed using partial least squares regression, after removing low signal-to-noise ratio wavelengths, and validated through a leave-one-out cross-validation procedure. In addition, the accuracy of the predicted values within each fraction was checked for each approach. Dividing the data set into subsets improved prediction models for each trait and for the samples in each milk fraction. Without considering milk fraction, the best improvement was observed for glucose and galactose. Glucose ratio performance deviation in cross-validation (RPD) increased from 7.42 to 11.31 and 11.06, for cluster and level approaches, respectively, whereas galactose RPD increased from 8.86 to 11.69 and 11.27 for cluster and level approaches, respectively. Considering milk fractions, the best improvement was observed for protein content, where RPD ranged from 0.08 to 6.06 for the whole data set calibration, whereas it ranged from 0.43 to 40.34 for the subset calibration approaches. Cluster and level approaches to build calibration models were comparable for samples from different fractions, suggesting that the 2 subsetting protocols should be both investigated to get the best prediction performances.