The China Meteorological Administration (CMA) has begun using a global artificial intelligence-based forecasting system nationwide. The system, named Fengqing, is a global short- and medium-range weather forecasting system based on artificial intelligence and is designated as CMA-AIM-GFS-Fengqing-ENS-V1.0.
According to Chinese publications in China Science Daily and IT Home, this system recently passed an operational test conducted at the National Meteorological Center.
It is important to note that the new system differs from the original Fengqing model. The earlier version, developed by the National Meteorological Center in collaboration with Tsinghua University in 2024, is a deterministic model that produces a single forecast run. This model contains approximately 3 billion parameters and can generate global forecasts for 15 days with a resolution of 25 kilometers in about three minutes.
The new version, however, functions as an ensemble system because it creates multiple forecast members to describe the uncertainty of atmospheric movements. This forms the basis for probabilistic forecasts of heavy rainfall, typhoons, and intense convection.
System Development and Technical Features
The system's development was led by the National Meteorological Center in cooperation with the Xunyan Institute of Artificial Intelligence Innovation for Meteorology, the National Meteorological Information Center, and Sugon. It is reported that the model combines the Vision Transformer architecture with physical constraints derived from meteorology, and all its data, algorithms, and computing power were sourced domestically.
Training and iterations were carried out on an AI cluster for science, which comprises about 100,000 accelerator cards located at the National Supercomputing Center in Tianjin.
Performance Results
CMA presents the results of an independent verification on samples covering the period from January 2025 to August 2026. The effective forecasting range for the mean global geopotential height at the 500 hPa level reached 11.0 days. Developers claim that key ensemble metrics, including the continuous ranked probability score of members and the error-spread ratio, showed good results. Some indicators surpassed those of the European Centre for Medium-Range Weather Forecasts' numerical ensemble and its AI-based system, AIFS-CRPS. It should be noted that these comparisons are proprietary to the developers, and detailed evaluations have not been published.
Like other AI-based forecasting models, this system is capable of completing a global calculation in mere minutes after training. CMA stated that the system will be used for disaster prevention, transportation, energy distribution, and environmental protection, and the team will continue working to improve forecasting accuracy and lead time.