Abstract In order to solve the problem of insufficient accuracy of the GPT3 model in simulating atmospheric profile, this paper analyzes its vertical correction method and then proposes an improved GPT3 model, namely GPT3v model, by introducing new temperature decline rate and pressure vertical correction algorithm. The new model is validated using sounding data, NCEP reanalysis data and GNSS data validation. The results show that in the vertical direction, the average RMSE values of temperature, pressure, and total zenith delay estimated by the GPT3v model are 6.0 K, 7.9 hPa, and 23.0 mm, respectively. Its accuracy is improved by 46%, 58%, and 52%, respectively, compared with the GPT3 model’s 11.1 K, 19.0 hPa, and 47.8 mm. The compared result of GNSS ZTDs shows that when there is a significant height difference between the station height and the surrounding four model grid points, the GPT3 model has a significant tropospheric delay estimation error, and the GPT3v model effectively overcomes this problem. These results demonstrate that the GPT3v model is superior to GPT3 in high-altitude and complex terrain areas.