Fengqing V1.0 passed its operational review and formally entered nationwide meteorological service on October 17, 2026. The announcement came from the National Supercomputing Internet at the 2026 National High Performance Computing Academic Conference: the AI global short-to-medium-range ensemble forecasting system is led by China's National Meteorological Center, jointly developed with the Xiong'an Meteorological AI Innovation Research Institute, the National Meteorological Information Center and the National Supercomputing Internet, and is China's first operational-grade AI ensemble weather model whose data, model and computing power are fully self-controlled.
An ensemble no longer reports a single line
The key to V1.0 is the word "ensemble." The first-generation Fengqing, which entered operational application in 2024, produced one deterministic forecast: at a given time and place, a specific temperature and a specific amount of rain. An ensemble instead perturbs the initial conditions and parameters slightly and runs dozens of members at once, then reads the distribution of the outcomes. When members agree closely, confidence in that weather event is high; when they split early, the system flags where the uncertainty sits. For questions like typhoon tracks, what forecasters actually need is never a lone line, but where the storm is most likely to go and how far it could plausibly stray.
An 11-day effective range, trained on a 100,000-card cluster
According to figures released by the developers, Fengqing V1.0 reaches a global effective forecast range of 11 days, with overall performance said to surpass the European Centre for Medium-Range Weather Forecasts' (ECMWF) AI ensemble forecasting system. The system was trained entirely on a 100,000-card AI4S computing cluster at a core node of the National Supercomputing Internet — the basis for the "self-controlled computing" claim: training data, model architecture and the cluster it runs on all sit inside the domestic system. It should be noted that the "surpasses ECMWF" statement is currently the developers' own wording; a metric-by-metric verification report has not been published, so for now it is recorded as the developers' claim.
Operational service means being graded every day
The Fengqing name is not new inside the meteorological system. The first generation, developed in 2024 by the China Meteorological Administration together with Tsinghua University, is already in operational use, and an evaluation published by the National Meteorological Center shows its 500 hPa geopotential height forecasts staying effective beyond 10 days, typhoon track forecasts slightly better than ECMWF's conventional IFS model, and a marked improvement in rainstorm TS scores at medium ranges — while typhoon intensity forecasts remained on the weak side. V1.0's switch to an ensemble architecture targets exactly those old problems: tracks roughly right, intensity off, and uncertainty left unstated. Having passed its operational review, it now faces the same daily verification as conventional numerical models; the thing to watch next is whether, during typhoon and flood seasons, the meteorological authorities publish head-to-head data between V1.0 and the ECMWF ensemble.
