Traditional tea blending which is heavily dependent on human sensory experience, is inherently time-consuming, labor-intensive, and subject to considerable subjective bias. These limitations make it increasingly inadequate to address the escalating market demand for rapid adaptation to multi-category teas flavor innovation, especially in the context of the current upgrading of tea consumption patterns. Thus, a digital blending model based on near-infrared spectroscopy and gaussian process regression (GPR) had been developed. A total of 886 tea samples from major production regions across China were used to establish a near-infrared quantitative model with six quality indicators via partial least squares method, which includes moisture, water extracts, tea polyphenols, caffeine, crude fiber and free amino acids. The calibration correlation coefficient of the model was higher than 0.92, and the external validation R-Square (R²) was greater than 0.98. Furthermore, 27 raw tea materials of different categories and shapes were selected to design blending schemes covering various scenarios. With near-infrared data as input, a GPR model combining radial basis function and white noise kernels was trained for quality prediction, which achieved an R² of 0.95, a mean square error of 0.12 on the independent test set, and a five-fold cross-validation R² of 0.92±0.04. The results of actual tests on several blended tea samples using national standard methods indicated that, despite the content of tea polyphenols, other five tea chemical components tested by the national standard methods basically fell within the prediction intervals provided by GPR model. Moreover, the noise index was the lowest in the complex scenario spanning multiple tea categories, with an overall mean noise of 2.79. This study provides technical support for transforming tea blending from an experience-driven approach to a data-driven one.