The crude fat content in fish feeds was determined using mid-infrared transmittance spectroscopy and chemometrics fast and non-destructively. A total of 225 samples were prepared for spectra collecting from a FT/IR-4000 Fourier Transform Infrared Spectrometer (400-4000cm-1). Principal component analysis (PCA) was carried out and spectral data were compressed into several new variables, which can explain the most variance of original spectra. The first six PCs were used as inputs of back-propagation neural network (BPNN) and least squares-support vector machine (LS-SVM) to create the calibration models. Compared with BPNN, a slightly better prediction precision was achieved based on LS-SVM with correlation coefficient (R) = 0.9757 and root mean square error for prediction (RMSEP) = 0.2579. The overall results indicated that mid-infrared spectroscopy incorporated to chemometrics was promising for the accurate assessment of crude fat content in fish feeds.