The grade quality of flue-cured tobacco plays an important role in the formulation design and the stability of the cigarette industry. In this paper, 768 tobacco samples from 40 prefecture-level cities in China, 2018 are selected for the experiment. The samples were classified and graded by traditional industrial grading method, including 7 different grade grades of tobacco leaves. The way to establish a tobacco quality grade prediction model by near-infrared spectroscopy and the near-infrared absorption spectrum characteristics of chemical groups and related components in different grades of tobacco are studied. The results show that the national tobacco grades prediction model is established in the non-segregated area, and the prediction standard deviation between the modeling set and the test set is not more than 1.35. After the samples are divided into five major production areas, models of each production area are established, and the prediction standard deviation of the models built in each production area after the division is found to be lower than that of the national model. The model in the Southeast region, the Southwest region and the Huanghuai region increased greatly, and the standard deviation of the test set was no more than 1. 07. After the standard normal transform (SNV) pretreatment of the average spectrum of different quality grades tobacco samples, the analysis is performed based on the information of the organic groups and related substances absorbed by the near-infrared light in different frequency ranges. It is found that tobacco with better quality grades has the characteristics of lower cellulose content and higher sugar content such as starch. The tobacco with lower quality grades has the characteristics of higher cellulose content and lower sugar content such as starch. At the same time, the worst quality grade (the upper and lower) tobacco has the characteristics of higher protein content. The results show that the application of near-infrared spectroscopy can realize the rapid prediction of the quality level of tobacco leaves. The prediction deviation is basically between adjacent levels, which meets the actual application requirements, and the prediction accuracy can be further improved by establishing prediction models of different production areas. At the same time, different grades of tobacco have different absorption characteristics of groups mainly composed of cellulose, starch, sugars, and proteins, which is also the information basis for applying near-infrared spectroscopy to achieve rapid detection of tobacco quality grades. This has important practical significance for improving the tobacco leaf grading evaluation system, further optimizing the grading scheme, and providing more scientific method guidance and technical support for product quality and maintenance.