Accurate estimation of gross primary production (GPP) is critical for global carbon cycle research. The current mainstream approaches, i.e., process-based models, are limited by insufficient parameter representation and high computing costs, and the burgeoning machine learning methods still suffer from inadequate training samples and poor transferability when applied to GPP estimation. Therefore, in this paper, we propose a process model-guided transfer learning approach for global GPP estimation, taking the low-resolution (0.5°) estimates from the Biosphere-atmosphere Exchange Process Simulator (BEPS) model as the source domain and eddy covariance (EC) data as the target domain. After joint constraint from the two domains, relatively high-accuracy GPP estimation at a resolution of 0.05° can be achieved after downscaled pre-training and fine-tuning based on EC tower data. The results indicate that the proposed framework can significantly improve the accuracy of GPP estimation, compared to a direct machine learning method based on only EC tower data (ΔR2 = 0.05, ΔRMSE = −1.02 g C m-2month-1) and the original BEPS estimates (ΔR2 = 0.05, ΔRMSE = −14.14 g C m-2month-1). The results of the temporal validation and regional cross-validation also show consistent results, indicating the superior spatio-temporal expandability of the proposed method. Furthermore, when compared with other global GPP products, the new global GPP product built in this study can effectively correct the underestimation/overestimation in high/low GPP regions in the existing machine learning-based GPP products (e.g., FLUXCOM GPP), especially in the area near the equator, and shows higher consistency with solar-induced chlorophyll fluorescence (SIF)-based and model-based GPP products. In addition to the new global GPP product, the results of this study also prove the reliability of combining a process-based model and a machine learning model in GPP estimation.