Google DeepMind announced on August 8, 2026, that it was open-sourcing WeatherNext 2 and WeatherNext Cyclones, AI forecasting models designed to predict tropical-cyclone tracks, intensity and wind structure. Results published in *Nature* showed a lead-time gain exceeding 24 hours on average, according to the company.
DeepMind described the gain this way: WeatherNext’s three-day predictions matched the accuracy that previous systems delivered at two days. It characterized that improvement as roughly equivalent to a decade of progress in meteorological forecasting. Those are company claims based on its evaluation; the supplied evidence does not include the paper’s detailed tables or an independent operational assessment.
Cyclone prediction normally presents competing scale requirements. Large atmospheric currents steer a storm’s path and are represented by global models, while intensity depends on localized processes near the core that benefit from higher resolution. WeatherNext uses one model to cover global weather and cyclone-specific behaviour, seeking to bridge that division.
The system was trained end to end on nearly 20 terabytes of global atmospheric data and the IBTrACS archive of almost 5,000 historical storms. DeepMind evaluated it on cyclones from 2023 and 2024, comparing deterministic and probabilistic results with leading weather models. Starting with global conditions, it can iteratively project weather and detailed cyclone tracks as far as 15 days ahead.
For each active cyclone, the current workflow generates 1,000 possible scenarios. That ensemble produces localized probability maps for tropical-storm and hurricane-force winds, giving forecasters a range of outcomes rather than one fixed track. Collaboration included Google Research, the US National Hurricane Center, the Cooperative Institute for Research in the Atmosphere, the UK Met Office and other weather agencies.
DeepMind says the model was used during the 2025 hurricane season and helped the National Hurricane Center anticipate Hurricane Melissa’s rapid intensification and landfall in Jamaica. The company links that forecast to earlier warning and more preparation time, although the supplied excerpt does not isolate WeatherNext’s contribution from the broader forecasting process.
Open sourcing gives researchers and forecasting organizations access to the models used during hurricane operations. DeepMind also points to possible applications in renewable-energy planning and other weather-sensitive decisions. Availability alone does not guarantee that agencies have the computing resources, local data or expertise to integrate the system.
Cyclone forecasts remain probabilistic and should support, not replace, expert forecasters and official warnings. The reported improvement is material because every additional hour can aid preparation, but real-world value depends on reliability across regions, rare storm behaviour and operational conditions beyond the historical evaluation set. Open access should also enable broader testing.
