Google DeepMind releases WeatherNext weights for localized cyclone forecasting
The release includes WeatherNext 2 and WeatherNext Cyclones checkpoints, plus a lighter version for researchers and weather agencies with constrained compute.
By Ryan Merket · Published
Why it matters
Open weights let weather agencies and researchers adapt Google's forecasting models locally, while a lighter checkpoint broadens testing. Weather-data access and the models' experimental status still limit operational use.
Google DeepMind (@GoogleDeepMind) released the code and model weights for WeatherNext 2 and WeatherNext Cyclones on August 6, 2026, giving researchers and weather agencies a base they can run and adapt for specialized or localized forecasting. The release follows operational use of a WeatherNext Cyclones checkpoint during the 2025 Atlantic hurricane season.
The work carries a collective byline. The paper and repository describe the project as work by a broader Google DeepMind and Google Research team, with researchers from the Google teams, the National Hurricane Center, the Cooperative Institute for Research in the Atmosphere and the UK Met Office among the contributors. The available material does not assign individual technical roles.
The open release is the consequential part of the announcement. WeatherNext had already appeared in Google products and public forecast feeds. Researchers can now inspect the implementation, download the weights, reproduce published cyclone evaluations and adapt models for regional forecasting problems. That gives national weather services and university groups a route to build on Google's work without waiting for another product integration.
The repository packages multiple checkpoints
The WeatherNext repository includes several checkpoints rather than a single downloadable model. One WeatherNext Cyclones checkpoint ran live during the 2025 Atlantic hurricane season. Separate checkpoints reproduce the paper's 2023 and 2024 evaluations, while a WeatherNext 2 checkpoint provides the broader medium-range forecasting system.
The repository includes separate WeatherNext 2 and WeatherNext Cyclones checkpoints. Google describes the code and weights as freely available for academic work, operational forecasting and the development of specialized or localized models.
Reproducing a full training run requires ERA5 and operational HRES weather data, according to the repository documentation. Those datasets carry their own access requirements and terms, creating a separate constraint for groups seeking to retrain the models rather than use released checkpoints.
The release also includes a lighter single-accelerator version for researchers with constrained compute. Google does not disclose the full models' computational cost, inference speed or energy use, leaving agencies and labs to assess whether they can operate the larger checkpoints at useful scale.
Forecasting the storm and its uncertainty
WeatherNext Cyclones predicts a storm's track, intensity and wind structure from global atmospheric conditions, producing forecasts as far as 15 days ahead. Google DeepMind says the system can generate 1,000 possible scenarios for each cyclone, which forecasters can turn into probability maps for tropical-storm and hurricane-force winds.
Large ensembles matter because a single forecast suppresses the uncertainty around dangerous, low-frequency outcomes. One scenario may show a storm weakening offshore, while a smaller group may show rapid intensification before landfall. Giving human forecasters a wider distribution helps them assess those tail risks rather than treating one model run as a definitive answer.
According to Google DeepMind's August 6 research announcement, WeatherNext Cyclones was trained end-to-end on nearly 20 terabytes of atmospheric information and the IBTrACS database of almost 5,000 historical storms. The system combines global weather dynamics with expert-curated cyclone records, addressing a longstanding forecasting trade-off between models that capture large atmospheric currents and specialized regional systems that resolve the processes around a cyclone's core.
Google evaluated the system on historical cyclones from 2023 and 2024 in its initial study, while later performance charts accompanying its Hurricane Melissa account include 2025 data. The Nature paper reports an average lead-time advantage of at least one day for track, intensity and wind-radius forecasts against leading operational systems. Its three-day forecast reached roughly the accuracy earlier models delivered at two days. The evaluation was conducted by researchers who helped build the system, and the paper includes forecasters from the NHC, CIRA and the UK Met Office as co-authors.
Google's open model and hosted service strategy
WeatherNext entered operational testing before the weights were released. Google DeepMind says the system supported the NHC during Hurricane Melissa in October 2025, predicting rapid intensification and a Category 5 landfall in Jamaica five days ahead. Google's account of the forecast says WeatherNext assigned an 80% probability to that outcome five days before landfall, rising to nearly 100% three days ahead. Official warnings still came from government forecasters using WeatherNext alongside physics-based models, satellite observations and aircraft data.
Open weights extend that collaboration model beyond agencies already working directly with Google. A forecasting service could tune the system around a coastline, a renewable-energy operator could adapt it to regional wind conditions, and researchers could probe where the model fails on rare extremes. Google DeepMind explicitly identifies localized and specialized models as intended uses of the release.
Google also retains a hosted distribution route. WeatherNext forecast feeds are available through Earth Engine and BigQuery, and Google offers custom model inference through Vertex AI. Users can operate the released models themselves or consume Google's managed outputs, though either route requires access to the relevant computing and weather data.
Google DeepMind labels WeatherNext an experimental research project rather than an officially supported Google product. The repository warns users to assess the risks of relying on generated forecasts, and Google DeepMind says WeatherNext does not replace alerts issued by national meteorological agencies. Researchers now have access to the forecasting system, while responsibility for turning probabilities into public warnings remains with human forecasters.