To study rapid analysis of cholesterol concentration in human serum with no reagent, near-infrared spectroscopy was used. Applicability of analytical models was studied. Spectra of serum were measured by a FT-NIR spectrometer with 1, 2 and 6.5 mm optical path length respectively. Partial least-square (PLS) models were calibrated for cholesterol in combination, first overtone and second overtone spectral regions. Root mean square error of prediction (RMSEP) of these models is 0.15, 0.16 and 0.29 mmol·L-1, and mean percent error of prediction (MPEP) is 2.9%, 3.1% and 4.8%, respectively. To validate applicability of these models, other two groups of serum spectra were measured in one month after the models were calibrated. These two sample groups were calculated using calibrated models. Prediction result of 1mm model is the best. The result of this experiment indicates tha