知Yizhi Daily
← Back to this edition

World

Chinese Research Team Releases Wantuo Meteorological Large Model, Capable of Improving Global Forecast Resolution at Low Cost

Edition 2026-11-17 · 08:00 Beijing time

The Institute of Tibetan Plateau Research of the Chinese Academy of Sciences released the Wantuo meteorological large model in Beijing on November 16. By 'freezing the parent model and training only increments,' the framework refined a 0.7° forecast model into a global two-level forecast system at 0.07° with updates every two hours in about four weeks on one 80GB-memory graphics card. Evaluations show improvements in forecast accuracy and typhoon-track performance.

The Institute of Tibetan Plateau Research of the Chinese Academy of Sciences released the Wantuo (One2N) meteorological large model in Beijing on November 16. The model was led by researcher Tian Xiangjun and developed jointly with Beijing Information Science and Technology University, Peking University, the Institute of Atmospheric Physics of the Chinese Academy of Sciences, the National University of Defense Technology, the Shandong Academy of Sciences and other institutions. The related paper was recently published on the preprint platform arXiv.

According to the introduction, Wantuo starts from a frozen parent model, and each level of forecaster equals the previous-level forecaster plus that level's unique incremental operator. During training, the parent model's parameters are frozen, gradients do not pass through the parent model, and only the newly added residual modules are trained. This eliminates the need to store the internal activations, parameter gradients and optimizer states required for the parent model's backpropagation, reducing memory overhead. The parent model can be a numerical model, a data-driven model or a hybrid of the two.

Evaluation results provided by the research team show that in 10-day global forecasts conducted based on 120 paired sets from 2023, Wantuo did not lose forecast skill while increasing resolution. Compared with the 0.25° data-driven Pangu-Weather forecast model, Wantuo had a lower root-mean-square error in 93.46% of 1,300 verification targets. Tests on 84 tropical cyclones in 2023 showed that its median typhoon track errors from 24 to 168 hours were 16.5% to 66.9% lower than Pangu's, and the 168-hour median error fell from 1,024.8 kilometers to 339.4 kilometers.

R&D personnel said Wantuo's value lies in separating existing capabilities from newly added capabilities. In the future, it can be used to independently upgrade various components of Earth system models, lowering the threshold for system iteration. The team also plans to promote a combination of 'domestic chips plus domestic specialized AI' and develop customized 'kilometer-minute-level' physical-AI meteorological simulators for regions such as the Qinghai-Tibet Plateau.

Original sources

中新网国内 ↗
Back to this editionOpen interactive reader