To break the dilemma on judging hard seeds and soft seeds of licorice and other legume families nondestructively, a distinguishing model for the hardness of licorice single seed was tried to be built by near infrared reflectance spectroscopy with distinguished partial least squares(DPLS). A total of 244 licorice seeds were divided into three groups： calibration set (120 samples), validation set (60 samples) and prediction set (64 samples), and each group has the same number of hard seeds and soft seeds. To eliminate the human error as far as possible, a specially made sample cup was designed for spectrum acquisition. Then the locations of the seed and the fiber-optic probe were fixed during each spectrum acquisition process. The influences of different replicate time, different spectral region and different calibration samples on the identification rate were compared. The result indicated that four replicates could increase the identification rate of the model significantly, the identification rates of the model of four replicates in calibration, validation and prediction set samples were 95.83%, 95.00% and 96.88% respectively, while that of one replicate were 93.33%, 91.67% and 82.81% respectively. The model of the spectral region between 4 000 and 80 000 cm ^-1 was better than that of other regions, and the identification rate in calibration, validation and prediction set samples were 95.53%, 95. 94% and 94. 53% respectively. Even with different samples, the predication rates were all more than 90%. The identification rates of hard seed and soft seed in prediction set samples were 92.50% and 96.56% respectively. The prediction for seeds with different size and different color showed that this model was not suitable for bigger and smaller seeds, especially not for black seeds. NIR offered a new way to distinguish the hardness of licorice singe seed quickly, precisely and nondestructively, which will advance the study on the mechanism of hardness of crop seeds.