China's Weather Agency Puts AI Ensemble Forecast Into Operation

A forecast is usually one guess about tomorrow's weather. Since the China Meteorological Administration (CMA) put its new Fengqing system into national operational use, the country's forecasters can ask the same question many times over and see how the answers spread. The release, designated CMA-AIM-GFS-Fengqing-ENS-V1.0, recently passed an operational review at the National Meteorological Centre.

The model's predecessor is easy to confuse with it, and the difference matters. The first Fengqing, developed with Tsinghua University in 2024, produces a single forecast run: about 3 billion parameters, 15-day global forecasts at 25-kilometre resolution, roughly three minutes per run. The new version is an ensemble. It generates many forecast members that describe the uncertainty in atmospheric motion, which is the basis for probability forecasts of heavy rain, typhoons and severe convection. A single run tells you what might happen; an ensemble tells you how likely each outcome is.

The National Meteorological Centre led development with the Xiong'an Meteorological Artificial Intelligence Innovation Institute, the National Meteorological Information Centre and the computer maker Sugon. The model pairs a vision Transformer architecture, a design borrowed from image recognition, with constraints drawn from meteorological physics. Its data, algorithms and computing were all sourced domestically. Training and iteration ran on an AI-for-science cluster of about 100,000 accelerator cards at the National Supercomputing Center in Tianjin.

On performance, CMA cites independent verification on samples from January 2025 to August 2026. The effective forecast range of the ensemble mean for global 500 hPa geopotential height reached 11.0 days. The developers say the key ensemble measures, including the probabilistic accuracy of the members and the match between spread and error, performed well, and that some indicators beat both the ECMWF ensemble and its AI counterpart, AIFS-CRPS. These comparisons are the developers' own, and the detailed scores have not been published, so outside experts cannot yet weigh them.

What changes for people is the speed and the warning. Like other AI forecast models, the system can complete a global run in minutes once trained, and CMA says it will serve disaster prevention, transport, energy dispatch and ecological protection. For emergency managers, a probability of a typhoon's track or of heavy rain over a specific region is more useful than a single line on a map. CMA says the team will keep working on accuracy and lead time; how far those gains will go, and when, is not yet known.