Atmospheric delay is a critical factor affecting high-precision navigation and GNSS water vapor inversion. The terrain in northwest China is highly variable, and the current zenith total delay (ZTD) model in this region does not adequately consider elevation, which cannot meet the real-time high-precision positioning needs. Therefore, it is urgently necessary to establish a high-precision ZTD model for northwest China. The ERA5 data from the European Centre for Medium-Range Weather Forecasts (ECMWF) for 2015—2017 was used to establish a ZTD model (MZTD model) for northwest China that considers non-linear elevation correction. This model addressed the problems of current ZTD models in northwest China, such as neglecting non-linear elevation correction and seasonal variations. The 2018 ERA5 data not involved in the modeling and the 84 GNSS stations data from the Northwest China land state network were jointly used to verify the accuracy and applicability of the MZTD model and to compare the accuracy with the GPT3 model, which was widely used today. The MZTD model's accuracy was compared with the widely used GPT3 model. The results show that, with 2018 ERA5 data and northwest China land network station data as reference values, the root mean square error (RMS) of the MZTD model is 3.14 cm and 2.81 cm, respectively, which is about 16.3% and 21.7% higher in accuracy than the GPT3 model. Moreover, the MZTD model has fewer model parameters, improving computational efficiency. Therefore, the MZTD model that considers non-linear elevation correction demonstrates better accuracy and applicability in the northwest China region, providing important references for real-time GNSS water vapor detection and navigation positioning in this area.