GenCast

GenCast, unveiled by Google DeepMind in December 2024 with a paper in Nature, is a diffusion-based ensemble weather forecasting model that generates 50 or more forecasts per run, each representing a possible weather trajectory. Trained on 40 years of historical weather data (1979–2018), it outperformed the European Centre for Medium-Range Weather Forecasts' ENS — the world's top operational system — 97.2% of the time on 2019 data, with particular strength in extreme weather and tropical cyclone tracks. A 15-day ensemble forecast runs in about 8 minutes on a single TPU, versus hours on a supercomputer for ENS.

Limitations: a research model rather than an open-weights release — Google publishes its forecasts rather than the model itself. A 2025 study also flagged that GenCast under-reproduces the atmospheric 'butterfly effect' relative to physics-based models, a reminder that skill scores alone don't capture physical fidelity.