Firefly algorithm (FA) combined with partial least squares (PLS) are developed for near infrared (NIR) spectral interval selection and quantitative analysis of complex samples. The method firstly segments the near-infrared spectra into a number of intervals. Vectors with 1 and 0, which represent the interval selected or not, are used as the inputs of the FA. The RMSEP value predicted by PLS model is used as the fitness function of the FA. The number of spectral intervals, the population number, environmental absorbance and the constant of FA are optimized. With the optimal parameters, FA-PLS model is established and applied to predict protein, hemoglobin and cetane number in wheat, blood and diesel fuel samples, respectively. The results show that FA-PLS can significantly improve the prediction accuracy compared with full-spectrum PLS model.