Google DeepMind beats top weather supercomputers with GenCast AI

GenCast is an open-source AI ensemble model from Google DeepMind designed for medium-range global weather forecasting up to 15 days ahead at 0.25° resolution.

Unlike traditional deterministic AI models that produce a single forecast, GenCast is a diffusion model that outputs an ensemble of 50 or more possible weather trajectories. This allows meteorologists and planners to accurately gauge probability, risk, and uncertainty.

What it does

Weather becomes chaotic after a few days, so weather agencies rely on ensemble forecasts: generating dozens of scenarios to see where the consensus lies. If 50 scenarios all track a hurricane making landfall in the same spot, confidence is high; if they scatter, uncertainty is large.

DeepMind evaluated GenCast against the European Centre for Medium-Range Weather Forecasts (ECMWF) ENS system across 1,320 targets combining different lead times and variables. GenCast outperformed ENS on 97.2% of all targets, and on 99.8% of forecasts with lead times beyond 36 hours.

It specifically excels at extreme event forecasting, including heatwaves, cold snaps, high wind speeds, and tropical cyclone tracks. In a proof-of-principle test predicting global wind farm power output, GenCast also beat ENS.

Under the hood and specs

GenCast adapts diffusion architectures to the spherical geometry of Earth. Given the current weather state, it learns to sample future weather states directly from complex probability distributions.

The model was trained on 40 years of ECMWF ERA5 reanalysis data covering variables like temperature, wind speed, and atmospheric pressure at various altitudes up to 2018, and validated on 2019 data.

Its hardware footprint is remarkably light compared to traditional numerical weather prediction. While physical systems require hours on supercomputers packing tens of thousands of processors, GenCast generates a single 15-day forecast in just 8 minutes on a single Google Cloud TPU v5, and each trajectory can be run in parallel.

DeepMind published the paper in Nature and made both the model code and weights open-source, with plans to release real-time and historical forecast data soon.

Why it matters

If you work in disaster preparedness, energy trading, or climate tech, GenCast is an essential open-weight model to explore. It brings ensemble forecasting—previously reserved for national meteorological supercomputers—down to standard cloud hardware with massive speedups.

Keep in mind that it does not replace traditional weather services entirely. GenCast still needs high-quality initial conditions and physics-based assimilation data from agencies like ECMWF. But by combining generative diffusion with climate science, DeepMind just proved that probabilistic AI forecasting is ready for real operational duty.