The protein and its component content in peanut seed is an important quality trait of peanuts. Exploring non-destructive and efficient content detection methods is an important research direction of peanut breeding and production. Constructing near-infrared models according to the color of the samples appearance is beneficial for improving detection accuracy. This study used 282 peanut germplasms with black, red, and pink coats to detect protein content by the Bradford method and the near-infrared spectral value. The Partial Least Squares Regression (PLSR) method was used to construct the near-infrared prediction model. A total of 11 near-infrared prediction models were constructed, including black seed coat crude protein, black seed coat albumin, black seed coat reaching, black seed coat contracting, red seed coat crude protein, red seed coat albumin, red seed coat reaching, red seed coat contracting, pink seed coat albumin, pink seed coat reaching, pink seed coat contracting, etc. The spectral value preprocessing method was a variety of composite processing methods. The best pretreatment methods for black seed coat crude protein, black seed coat albumin, black seed coat reaching, and black seed coat contracting models were Baseline + Detrend, Detrend + MSC, 2nd-der+Detrend+1st-der, Baseline + SNV, respectively. The best pretreatment methods of red seed coat crude protein, red seed coat albumin, red seed coat reaching, and red seed coat contracting model were Baseline + SNV, Baseline + SNV + MSC, SNV + MSC + Baseline, SNV + MSC + Baseline, respectively. The best pretreatment methods of pink seed coat albumin, pink seed coat reaching, and pink seed coat contracting model were 2nd-der+1st-der, 2nd-der+Detrend+1st-der and 2nd-der+Baseline+1st-der, respectively. The models correlation coefficient (Rc) was 0.825~0.925, and the root means standard error of calibration (RMSEC) was 0.110 ~1.383. The correlation coefficient (Rp) of the external validation set of the 11 models ranged from 0.822 to 0.971, and the root mean standard error of prediction (RMSEP) ranged from 0.102 to 0.954. The peanuts with different seed coat colors were detected by other color models and fitted with their chemical values. The correlation coefficients were in the range of 0.002~0.877, and the standard errors were in the range of 0.257~9.464. The correlation coefficients were lower than the correlation coefficients of the external validation set, and the best detection model was the model corresponding to the seed coat color. In this study, a model of peanut protein and its component content with different seed coat colors was constructed, which can quickly and non-destructively detect the content of peanut protein and provide the basis for the selection of raw materials for peanut protein processing and the breeding of specials peanut germplasm.