High spatial resolution Gross Primary Productivity (GPP) estimation makes it feasible to better understand the spatial heterogeneity of mountain vegetation photosynthesis. Spatial downscaling is a practical approach to obtaining high resolution GPP estimates, whereas there is almost no attention given to the downscaled biases resulted from the widely reported scaling errors in medium or coarse resolution GPP estimates. To fill the above gap, this study adopted an eco-hydrological model to obtain 960 m resolution distributed GPP estimates (not including scaling errors) and lumped GPP estimates (including scaling errors) over four mountainous watersheds. Then, the distributed and lumped estimates were downscaled from 960 m to 30 m, respectively. Finally, the contribution of reducing scaling errors was characterized by the agreement index (d), BIAS and Root-Mean-Square-Error (RMSE) values between downscaled GPP and referenced GPP (directly generated at the spatial resolution of 30 m). Results showed that a large difference existed between lumped and distributed GPP, with d, BIAS, and RMSE of 0.79, 212, and 334 gC m-2 year-1, demonstrating that the scaling errors should be given enough attention to current coarse resolution GPP estimates. Before considering the scaling errors, large uncertainties were observed in the GPP downscaled from lumped values, with d, BIAS, and RMSE of 0.68, 220, and 480 gC m-2 year-1. After considering the scaling errors, a significant improvement was achieved in the GPP downscaled from distributed values, with an increased d value of 0.81, a decreased BIAS value of 10 gC m-2 year-1, and a decreased RMSE value of 388 gC m-2 year-1, indicating that reducing the medium or coarse resolution scaling errors would effectively improve the spatial downscaling of mountain vegetation GPP. Our study highlights the effect of scaling errors on the spatial downscaling of mountain vegetation productivity, which should be given more attention in the future carbon modeling.