Google DeepMind open-sources WeatherNext after a major advance in AI cyclone forecasting

Google DeepMind published a significant weather-AI research update on 6 August 2026: its WeatherNext Cyclones model is designed to forecast a tropical cyclone's track, intensity and wind structure together, and the company is now open-sourcing the models used in the work.

What changed

According to Google DeepMind, WeatherNext Cyclones achieved state-of-the-art results across cyclone track, intensity and wind-structure forecasting on the evaluation described in its accompanying research. On average, the model's three-day forecasts reached accuracy comparable to what prior leading systems achieved at about two days, giving forecasters roughly one additional day of predictive lead time.

The system combines global weather modelling with specialised historical cyclone observations. Rather than producing only one deterministic forecast, it can generate large ensembles of possible futures so forecasters can examine uncertainty and lower-probability but high-impact scenarios such as rapid intensification. DeepMind says the current system can generate 1,000 scenarios for a cyclone and forecast out to 15 days.

Why the open-source release matters

DeepMind is releasing WeatherNext Cyclones, WeatherNext 2, code and model weights for research and operational experimentation. It also describes WeatherNext 2-mini, a smaller version that can run on a single TPU through a public Colab notebook. That creates a practical path for academic groups, meteorological agencies and nonprofits to test specialised or local forecasting workflows without rebuilding the full system from scratch.

The research is especially notable because cyclone forecasting has historically involved a trade-off between large global models that capture storm tracks and high-resolution local models that better represent intensity. WeatherNext is intended to address both within one AI forecasting approach.

Real-world context

Google DeepMind says an earlier WeatherNext system supported the US National Hurricane Center during the 2025 hurricane season, including forecasts that helped anticipate Hurricane Melissa's rapid intensification and landfall in Jamaica. The 2026 research release extends that work and formalises the model's evaluation across historical cyclones.

The practical significance is not that AI replaces official forecasting agencies: DeepMind explicitly says WeatherNext outputs are experimental and that people should continue to use their national meteorological agency for official warnings. The important development is that AI weather models are becoming more capable, faster to run and increasingly available to the broader scientific community.

For researchers and students, the open-source release may also create new opportunities around climate science, machine learning, geospatial analysis, extreme-weather modelling and public-sector forecasting infrastructure.