Gross primary productivity (GPP) is an important component of the global ecosystem carbon cycle and has a significant impact on global climate change. Accurate estimation of GPP helps to understand the carbon cycle processes in atmospheric and terrestrial ecosystems. a variety of remote sensing models are available to simulate GPP, but few studies have assessed the applicability of remote sensing GPP models for arid and semi-arid regions, and semi-arid ecosystems dominate the trend and inter-annual variability in the land CO₂ sink, it is important to compare the applicability of different remote sensing models in the arid zone of Central Asia to advance the estimation of carbon balance in the global arid zone. In this study, the data from four ground stations observed based on the eddy covariance technique are used as validation data to verify the simulation accuracy of five models, MOD17, VODCA2, VPM, TG, and SANIRv. The four ground stations are China Fukang Desert Ecological Observation and Research Station (CN-Fuk), China Ulan-Usu Oasis Farmland Ecology and Agrometeorological Experiment Station (CN-Wul), Kazakhstan Desert Ecological Observatory (KZ-Ara), and Kazakhstan Lake Balkhash Grassland Ecological Observatory (KZ-Bal). The remote sensing model is mainly driven by moderate-resolution imaging spectroradiometer (Modis) remote sensing data and ERA5-land meteorological reanalysis data. The results indicated that:(1) MOD17 and VPM models based on light energy utilization theory simulated the highest accuracy of GPP in Aral Sea and Fukang desert vegetation (R² 0.52 and 0.80, respectively), but there was more obvious underestimation (RE>20%) in simulating the productivity of grassland and farmland ecosystems. The remote sensing models based on vegetation index, TG model, SANIRv model, simulating Grassland ecosystem of Lake Balkhash and farmland ecosystem of Ulan-Usu had the highest accuracy (R² 0.91 and 0.81, respectively), and the relative error between simulated and measured values was low. The microwave-based VODCA2 model simulated the productivity of each ecosystem with the worst effect, also the VODCA2 model performs the worst in tracking the seasonal dynamics of GPP in four ecosystems, and the other four remote sensing models show excellent performance. (2) Water deficit was the main factor limiting vegetation GPP, so whether water stress is reasonably considered is an important factor affecting the applicability of GPP models in the arid zone of Central Asia. This study initially reveals the application potential of remote sensing GPP models in the arid zone of Central Asia, which can help to advance the accurate estimation of global vegetation carbon fluxes.