DeepMind open-sources a cyclone AI model

Google DeepMind published research in Nature on August 6 showing state-of-the-art accuracy for tropical cyclone track, intensity, and wind-structure forecasts, then open-sourced the code and model weights behind it under an Apache 2.0 license. Three variants are available: WeatherNext Cyclones, tuned specifically for storm forecasting; WeatherNext 2, the general-purpose successor model; and WeatherNext 2-mini, a lighter version DeepMind says can run inside a free Google Colab notebook.

The headline number is lead time. The models extend reliable cyclone-track warning by roughly 24 hours, so a three-day forecast now carries the accuracy that previously required waiting until two days out. WeatherNext Cyclones runs on a coarser 28×28 km grid (111×111 km for the mini variant), trading some resolution for speed. DeepMind says the system was already tested operationally during the 2025 Atlantic hurricane season, when the U.S. National Hurricane Center used it to help forecast Hurricane Melissa’s rapid intensification and landfall in Jamaica.

What it means for you

Unlike most “open” AI announcements this year, this one ships weights and code together under a genuinely permissive license, not a gated research preview, a contrast to the narrower open-weight positioning most labs, including Anthropic, have taken. If you build anything downstream of weather data (insurance risk models, logistics routing, agricultural planning, climate-adjacent products), WeatherNext 2-mini is worth a weekend to try given it fits on free-tier compute. It also lands the same week as DeepMind’s own leadership shakeup under Jeff Dean, a reminder that the org’s public research output hasn’t slowed even as its structure changes.