The miniaturization of near-infrared spectrometers has been growing rapidly. Several designs are now available, but there is a lack of understanding of how spectral data from these designs are affected by complex matrices and what are the limitations when compared to established systems. This study compares a popular miniaturized NIR device based on Hadamard-transform spectrometer (named miniaturized NIR) with a system based on dispersive spectrometer (named handheld-NIR) to assess: 1) their predictive performance; 2) the effect of a complex matrix on the performance, and 3) ability to discriminate multiples compounds in that matrix. The devices were challenged with a wide range of cheese types (n = 36) from different species (cow, goat, ewe and buffalo), brands (n = 30), countries of origin (n = 9) and with a broad range of cheese matrices (soft, fresh, semi-hard, hard and aged) to predict fat composition. Spectra were collected non-invasively with no sample preparation. Three wavelength ranges from handheld NIR were compared to miniaturized NIR based on two modelling approaches were used: a linear (Partial Least Square - PLS) and a non-linear (Support Vector Machine - SVM). The important wavelengths for each model were identified and used to assess the ability of the spectral data to differentiate among fatty acids. The highest prediction performance was observed for saturated fatty acids (C4.0, C14.0, C15.0 C16.0, total SCF and total SFA) with the RPDEXT-VAL for the external validation dataset presenting values higher than 3 and the coefficient of determination for the external validation dataset (R2EXT-VAL) higher than 0.89, mostly for SVM models. The sum of fatty acids also shows good prediction performance with RPDEXT-VAL higher than 3 and R2EXT-VAL higher than 0.89. Models with RPDEXT-VAL between 2 and 3 includes: C6.0; C17.0; C18.0; C10.1; C16.1; C17.1; iso.C15.0; iso.C.16; iso.C17; C18.1.c11; C18.1.c9; anteiso C17; total MUFA; and total BCFA. The cheese matrix affected the linearity between spectral data and fatty acids concentration requiring a more complex model (SVM), but this effect was not enhanced by the instrument type. It was shown that the spectral information allows discrimination among fatty acids and this ability was not affected by the type of instrument. These findings demonstrated that the miniaturized NIR can be directly applied to a cheese matrix to monitor fatty acid composition with results equivalent to an optical-based design.