Radial basis function neural network (RBFNN) combining with near infrared spectroscopy (NIRS) is applied to develop quantitative analyzing models of mannitol, polysaccharide and adenosine in Marasmius androsaceus fermentation mycelium. Using submerge fermentation, 164 Marasmius androsaceus mycelium samples are obtained. The contents of mannitol, polysaccharide and adenosine are determined via traditional methods and the near infrared spectroscopy data of the 164 samples are collected. The outliers are removed and the number of calibration set is confirmed via Monte Carlo partial least square (MCPLS) method. Based on the values of degree of approach (Da), the moving window radial basis function neural network (MWRBFNN) is applied to optimize characteristic wavelength variables, pre-processing methods, hidden layer nodes (NH) and spreads in the models. The quantitative analyzing models of mannitol, polysaccharide and adenosine in Marasmius androsaceus fermentation mycelium are developed successfully. The correlation coefficients between the reference values and predictive values of mannitol, polysaccharide and adenosine in both of the calibration set and validation set of optimum RBFNN-NIRS models are 0.9274, 0.9009, 0.9440 and 0.9354, 0.9018, 0.8847 respectively. All the data suggest that these models possess excellent fitness and predictive ability.