DeepMind Open-Sources WeatherNext Cyclones After Nature Paper — 3-Day Forecasts Now as Accurate as Prior Models' 2-Day, ~100 km Position Error
Google DeepMind / Hacker News·high signal
Google DeepMind published WeatherNext Cyclones in Nature on August 6, 2026, reporting roughly 100 km position error on 3-day tracks and ~11 knot intensity error, beating ECMWF-ENS on track and HWRF on intensity across 2023–2025 evaluations. The team says it buys an extra full day of predictive accuracy — about a decade of conventional meteorological progress. Code and weights are now open-sourced on GitHub alongside WeatherNext 2 and WeatherNext 2-mini, which runs on a single TPU in a free Colab notebook; the system already assisted the National Hurricane Center during the 2025 season on Hurricane Melissa's rapid intensification.