BEAST: Exascale Bayesian Weather Transformer Trained on JUPITER Reaches 3.96 Exaflops

A multi-institution research team says it has trained a Bayesian weather-forecasting transformer at exascale on Europe’s JUPITER supercomputer, reaching a reported 3.96 exaflops while producing global forecasts with quantified uncertainty at 0.25-degree resolution. In a new arXiv preprint, the researchers say the system, called BEAST, can deliver probabilistic forecasts that are competitive with leading AI and numerical weather systems.

That matters because weather forecasting is not just about the single most likely outcome. Forecasters also need a measure of uncertainty — how confident a model is, and how widely possible outcomes may vary — especially when tracking extreme events. The new work, posted publicly as arXiv:2609.12815v1 on Sept. 11, 2026, has not been peer reviewed, and the results have not been independently audited in the source material.

The paper, titled “4D Parallelism Unlocks Exascale Bayesian Neural Networks for High-Fidelity Atmospheric Modeling,” comes from researchers affiliated with institutions including Karlsruhe Institute of Technology, Forschungszentrum Jülich, CSC – IT Center for Science in Finland, and NVIDIA. In the abstract, the authors write: “We present BEAST, the first-ever Bayesian Swin Transformer for atmospheric forecasting on 0.25° global resolution able to accurately quantify both aleatoric and epistemic uncertainty.” In plain terms, the model is designed to estimate both uncertainty arising from the atmosphere’s inherent randomness and uncertainty tied to the model itself. The full training run described in the paper used a 700 million-parameter version trained on 40 years of ERA5 reanalysis data, covering 1979 through 2020, with 983,350 optimizer update steps. The authors say one epoch over that dataset required 89,835 gradient update steps and took 25.45 hours.

The core technical claim is less about a new weather architecture than about scale. The team says it built a 4D parallelization scheme — combining domain-tensor parallelism, uncertainty parallelism and data parallelism — to make large Bayesian neural networks train efficiently across thousands of GPUs. For a larger, 2.4 billion-parameter configuration, the paper reports a peak performance of 3.96 exaflops on 20,480 NVIDIA GH200 GPUs on JUPITER, with 73.99% scaling efficiency.

On forecast quality, the preprint says BEAST achieved competitive probabilistic skill against Functional Generative Networks, or FGN, and the European Centre for Medium-Range Weather Forecasts’ IFS ensemble, known as IFS-ENS. The reported comparisons cover metrics including root mean square error and continuous ranked probability score for near-surface wind and temperature, as well as upper-air wind and temperature. The paper also says the model showed strong skill in extreme-event case studies, including Europe’s 2022 heat and freeze events, though those findings have not been independently verified.

The authors also make aggressive speed claims at inference, the stage when a trained model produces forecasts. They report that a single 15-day forecast trajectory takes about 16 to 17 seconds on four GH200 GPUs. At larger scale, the paper says the system reached 14.08 forecasts per second with 96 ensemble members on 1,536 nodes. In the abstract, the authors say this lets BEAST “generate large ensembles 3 to 4 times faster than the current-best AI model.” But the paper’s own comparison is to FGN running on a TPU v5p, not the same hardware. The authors note in the paper body: “For comparison, FGN reports inference times of just under 1 minute on a single TPU v5p.” That makes the speed comparison directionally interesting, but not an apples-to-apples benchmark.

JUPITER, an exascale-class supercomputer at Forschungszentrum Jülich in Germany, is built around NVIDIA GH200 Grace Hopper systems, making it a natural test bed for very large AI and high-performance computing workloads. The broader significance of the preprint is that it argues uncertainty-aware weather AI can be trained at that scale, not simply that AI has surpassed traditional forecasting. Still, key reproducibility pieces are missing: the arXiv version does not link a public BEAST code repository, and the source material reports no outside validation, independent benchmark audit or operational deployment.