Near infrared (NIR) spectroscopy calibration models have been developed for simultaneous and rapid determining the adenosine, protein, polysaccharide and cordyceps acid contents in cordyceps militaris mycelium powder. Cordyceps militaris mycelium powder samples are collected by culturing different cordyceps militaris mutant strains under various cultured conditions and the NIR spectra are acquired. Radial basis function neural network (RBFNN) is used for modeling the relation between the NIR spectra and the adenosine, protein, polysaccharide and cordyceps acid contents. The RBFNN models have been optimized. Firstly, movable window method is employed to select characteristic wavelength variables. And then, spectra preprocessing methods, the number of hidden nodes and the spread constant are selected. The optimum RBFNN models for determining the adenosine, protein polysaccharide and cordyceps acid contents are obtained. The correlation coefficients (Rc) between reference values of calibration set and the predictive values are 0.9436, 0.9884, 0.9018 and 0.8848, respectively, which indicate that the fit of these models is satisfying. The root mean square error of prediction set (RMSEP) RMSEP are 0.6225 mg/g, 0.0179 g/g, 0.0115 g/g, and 0.0102 g/g, respectively, which indicate that the predictive capabilities of the models are fine.