Rapid prediction of pollutant dispersion after a nuclear accident is critical for nuclear emergency response. Atmospheric dispersion models (ADMs) are widely used due to their computational speed and accuracy balance. This study uses the GPU-accelerated QES (GPU-QES) model for conducting atmospheric dispersion modeling at both small-scale (3 km) and local-scale (15 km) areas for the same nuclear power plant (NPP) site. The model’s performance and accuracy at different scales were validated by comparing its simulation results with 2D horizontal and vertical wind tunnel measurements. The results show that GPU-QES generates 3D wind fields for both small and local scales across the entire domain in under 5 s, with smooth plumes produced in 10 min for the small-scale and 30 min for the local-scale simulations. The small-scale simulation accurately captures vortex winds and plume details between buildings with a 5 m horizontal grid. In contrast, the local-scale simulation shows strong agreement with measurements in mountainous areas with a 50 m horizontal grid. Statistical analysis shows that the results are within the acceptable range at both scales. FAC2 reaches 0.55 at the local scale and 0.39 at the small scale, showing improvements of 17.6% and 15.6% over the Micro-SWIFT-SPRAY (MSS) model, respectively. With its high speed and accuracy, GPU-QES is highly advantageous for nuclear emergency response.