Near-infrared reflectance spectroscopy (NIRS) (700–2500 nm) was used to predict milk fatty acid (FA) composition. Broad FA variability was ensured by using experimental cow milk derived from different feeding regimes (pasture and preserved forages with or without lipid supplements). Detailed FA composition was analyzed by gas chromatography. Predictive equations (354 samples) were developed for liquid and oven-dried milk samples using modified partial least squares with cross-validation and external validation (114 samples). Coefficient of determination in external validation (R2V) and residual predictive deviation (RPD) were good (R2V ≥ 0.88; RPD ≥ 3.26) for saturated fatty acids (SFA), monounsaturated fatty acids (MUFA), unsaturated fatty acids (UNSAT), trans FA, trans and cis-C18:1, caproic, caprilic, capric, lauric, myristic, palmitic and oleic acids in oven-dried milk, approximate for polyunsaturated fatty acid (PUFA), stearic, vaccenic and rumenic acids (R2V ≤ 0.81; RPD ≤ 3.23) and poor for linoleic, linolenic, total n-6 and n-3 acids. The quantification was more accurate for oven-dried milk, but good results were also obtained for SFA, MUFA, palmitic and oleic acids in liquid milk.