Where can I download WeatherNext?

WeatherNext Cyclones, WeatherNext 2, and the compact WeatherNext 2-mini are on GitHub at google-deepmind/weathernext under Apache-2.0. WeatherNext 2-mini runs on a single TPU in a free Colab notebook, so you can try inference without setting up hardware.

How much better is WeatherNext at cyclone forecasting?

DeepMind reports that WeatherNext Cyclones delivers an average of 24 hours or more of lead-time advantage over leading operational models for cyclone track, intensity, and wind structure. Put simply, its three-day forecast matches what prior systems needed four days to produce.

Is this the same model as the WeatherNext 2 released in November 2025?

No. WeatherNext Cyclones is a new, cyclone-focused variant whose results are the subject of the August 2026 Nature paper. WeatherNext 2 — the general-purpose forecasting model DeepMind operationalized in October 2025 — is being released alongside it, together with WeatherNext 2-mini for lightweight use.

Did WeatherNext ever help issue a real hurricane warning?

Yes. During the 2025 Atlantic season the model assisted the National Hurricane Center on Hurricane Melissa, predicting rapid intensification and the Jamaica landfall in time for the NHC to issue an advance warning. The paper cites this as the first operational use of an AI cyclone model at that agency.

Who evaluated WeatherNext before this release?

DeepMind worked with the National Hurricane Center, the Cooperative Institute for Research in the Atmosphere at Colorado State, the UK Met Office, and NOAA's National Weather Service and NCEP. The Nature paper compares WeatherNext against these agencies' operational models on storms from 2023 through 2025.

Google DeepMind · 2026-08-06 · major

WeatherNext Cyclones — DeepMind's Nature-published model adds a day of warning

WeatherNext Cyclones matches the National Hurricane Center on cyclone track, intensity, and wind structure with a day of extra lead time. DeepMind published the result in Nature and open-sourced three model variants on GitHub.

Illustrated tropical cyclone swirl labeled WeatherNext Cyclones on a blue background

A neural cyclone forecaster that matches operational models one day sooner and ships open under Apache-2.0.

Quick facts

Maker	Google DeepMind
Models released	WeatherNext Cyclones, WeatherNext 2, WeatherNext 2-mini
License	Apache-2.0
Publication	Nature (Aug 6, 2026)
Ensemble size	1,000 members
Resolution	28 x 28 km
Training data	~20 TB atmospheric reanalysis + ~5,000 historical storms (IBTrACS)
Evaluators	NHC, UK Met Office, CIRA, NOAA NWS/NCEP

What is it?

WeatherNext Cyclones is a diffusion-based ensemble model from Google DeepMind that predicts a tropical cyclone's track, intensity, and wind structure directly from atmospheric data. Alongside the paper, DeepMind released three variants — WeatherNext Cyclones, WeatherNext 2, and WeatherNext 2-mini — on GitHub under Apache-2.0. The mini version runs on a single TPU in a free Colab notebook.

How does it work?

Trained on roughly 20 terabytes of global atmospheric reanalysis data plus about 5,000 historical storms from the IBTrACS database, the model generates 1,000-member ensembles at 28x28 km resolution — a hundred times coarser than physics-based systems but faster and, per the Nature paper, more skillful on cyclone metrics. During the 2025 hurricane season it helped the National Hurricane Center flag Hurricane Melissa's rapid intensification and Jamaica landfall.

Why does it matter?

An extra 24 hours of warning is roughly a decade of meteorological progress delivered in a single release, according to DeepMind's evaluation. Forecasters at the NHC, the UK Met Office, CIRA and NOAA's NWS/NCEP evaluated the model, and the open-source release lets any research group reproduce or fine-tune it — a rare combination for a frontier weather system.

Who is it for?

climate scientists, meteorologists, disaster-response teams, AI-for-science researchers

Frequently asked questions

Where can I download WeatherNext?: WeatherNext Cyclones, WeatherNext 2, and the compact WeatherNext 2-mini are on GitHub at google-deepmind/weathernext under Apache-2.0. WeatherNext 2-mini runs on a single TPU in a free Colab notebook, so you can try inference without setting up hardware.
How much better is WeatherNext at cyclone forecasting?: DeepMind reports that WeatherNext Cyclones delivers an average of 24 hours or more of lead-time advantage over leading operational models for cyclone track, intensity, and wind structure. Put simply, its three-day forecast matches what prior systems needed four days to produce.
Is this the same model as the WeatherNext 2 released in November 2025?: No. WeatherNext Cyclones is a new, cyclone-focused variant whose results are the subject of the August 2026 Nature paper. WeatherNext 2 — the general-purpose forecasting model DeepMind operationalized in October 2025 — is being released alongside it, together with WeatherNext 2-mini for lightweight use.
Did WeatherNext ever help issue a real hurricane warning?: Yes. During the 2025 Atlantic season the model assisted the National Hurricane Center on Hurricane Melissa, predicting rapid intensification and the Jamaica landfall in time for the NHC to issue an advance warning. The paper cites this as the first operational use of an AI cyclone model at that agency.
Who evaluated WeatherNext before this release?: DeepMind worked with the National Hurricane Center, the Cooperative Institute for Research in the Atmosphere at Colorado State, the UK Met Office, and NOAA's National Weather Service and NCEP. The Nature paper compares WeatherNext against these agencies' operational models on storms from 2023 through 2025.

Try it

github.com/google-deepmind/weathernext