DeepMind and ECMWF release an AI weather model that matches supercomputer accuracy at a fraction of the cost.
A collaboration between Google DeepMind and the European Centre for Medium-Range Weather Forecasts (ECMWF) has produced GenCast-2, an AI weather prediction model that matches the accuracy of traditional numerical weather prediction systems while running 10 times faster on standard GPU hardware.
Traditional weather models solve complex fluid dynamics equations on massive supercomputer grids—a process that takes hours for a single 10-day forecast. GenCast-2 replaces this with a diffusion-based generative model trained on 40 years of global atmospheric data, producing ensemble forecasts in minutes on a cluster of eight A100 GPUs.
In head-to-head testing against ECMWF's flagship HRES model, GenCast-2 showed equal or better accuracy on temperature, wind speed, and precipitation forecasts out to 10 days. Beyond 10 days, the AI model's probabilistic ensemble approach provides more calibrated uncertainty estimates—crucial for disaster preparedness.
The implications extend far beyond weather. The same architecture is being adapted for long-range climate projections, ocean current modeling, and air quality forecasting. A variant trained on historical El Niño data successfully predicted the timing of the 2025–2026 event six months in advance.
For climate researchers and policy analysts, Vincony's Deep Research tool can synthesise published benchmarks, datasets, and methodological comparisons across AI and traditional climate models—accelerating literature reviews that would otherwise take weeks.
The model weights have been released under an open licence, allowing national meteorological services worldwide to deploy GenCast-2 locally. Several developing nations that lack supercomputer access are already piloting the system for regional weather forecasting.