The concentration of casein and serum protein and the relative proportion of casein to true protein in milk-based beverages and microfiltration (MF) retentates will influence their sensory and functional properties. Therefore, control of protein concentration and the ratio of casein and serum protein may be important commercially. Our objective was to develop partial least square (PLS) models using mid-infrared (MIR) spectra to predict true protein (TP), casein (CN), serum protein (SP), and CN as percentage of TP (CN%TP) content of microfiltration (MF) retentates and unflavored milk-based beverages. A total of 625 milk formulations varying in fat (2.5, 3.2, 3.9, 4.6, 5.3%), TP (2.5, 3.1, 3.7, 4.3, 4.9%), CN%TP (71, 75, 79, 83, 87%), and anhydrous lactose (3.5, 4.0, 4.5, 5.0, 5.5%) were produced using different combinations of skim milk ultrafiltration permeate, serum protein isolate, microfiltration retentate, cream, lactose monohydrate, and distilled water. MIR spectra were collected for each formulation, and in addition all formulations were analyzed in duplicate by Kjeldahl for total nitrogen, non-protein nitrogen, and non-casein nitrogen. Separate PLS models were developed for prediction of TP, CN, SP concentration (g/ 100 g of beverage) and CN%TP using the spectral ranges: (3,000 to 2,750, 1,800 to 1,700, and 1,580 to 1,000 cm⁻¹). The relative predictive differences (RPD) for TP, CN, SP, and CN%TP PLS models were 62.70, 31.45, 14.10, and 3.79, respectively, the standard errors of cross validation (SECV) were 0.0139, 0.0220, 0.0183, and 1.2405%, respectively, and R-squared for the models were 0.999, 0.999, 0.995, and 0.930, respectively. PLS models with RPD values <8 are not accurate enough for analytical purposes. The standard deviation of the difference (SDD) between reference and PLS predicted CN%TP was 1.236%. A more accurate prediction of CN%TP was achieved by using the PLS predicted CN and TP to calculate CN%TP (SDD = 0.558%).