$ cat wiki/models/weathernext-cyclones.md
WeatherNext Cyclones
Spec
| Attribute | Value |
|---|---|
| Developer | Google DeepMind |
| Released | 2026-08-06 |
| Announced | 2026-08-06 |
| Context window | unknown |
| Pricing | unknown |
| License | unknown |
| Availability | GitHub (code + weights), Colab notebook (2-mini variant) |
Context window and Pricing are unknown because neither applies in the form the row | |
| asks for and no source read supplies an equivalent: this is a weather model distributed as | |
weights, not an API sold on tokens. License is unknown as a genuine gap — the release | |
| is described as open-source in every source read, but no source names the licence | |
| (source). |
Release Date
2026-08-06, simultaneously with a Nature paper (source).
Benchmarks
Evaluated on cyclones from 2023 through 2025: WeatherNext's track, intensity and wind-structure predictions carry an average of a day or more of advantage over leading operational models — its three-day forecast matches what prior systems delivered at two days (source).
The model produces large ensembles of possible global weather and cyclone scenarios extending 15 days ahead (source).
The figure to hold onto is the lead time, not an accuracy percentage. Cyclone forecasting is one of the few domains where the decision the forecast feeds — evacuate, stage, or wait — is measured in hours, so a day of additional warning is the metric rather than a proxy for it.
Which operational baselines the "day or more" is measured against — ECMWF, the National Hurricane Center's official forecast, or both — is not established in any source read (source).
Use Cases
- Operational tropical-cyclone forecasting — track, intensity and size, worldwide. The Nature paper is titled Operational Tropical Cyclone Forecasting with AI, and the work was done with operational forecasters at the National Hurricane Center and the Cooperative Institute for Research in the Atmosphere (source).
- General global weather ensembles via the accompanying WeatherNext 2 release.
- Local experimentation — WeatherNext 2-mini runs on a single TPU in a free Colab notebook (source).
Compared To
Three variants were released together, and they are not interchangeable (source):
| Variant | What it is |
|---|---|
| WeatherNext Cyclones (WN-C) | ran during the hurricane season; the results in the Nature paper are its |
| WeatherNext 2 | the later general forecasting update, operationalised October 2025 |
| WeatherNext 2-mini | compact; single TPU, free Colab notebook |
| Within this wiki it sits alongside Google DeepMind's other science models — | |
| AlphaEvolve and Co-Scientist (Google DeepMind) — but differs from both in one respect | |
| that matters: the weights were released. AlphaEvolve and Co-Scientist were published as | |
| results; this was published as an artefact anyone can run. |
Open Questions
- What licence governs the released weights?
- Parameter counts and training compute for any of the three variants.
- Is WN-C reachable through the Gemini API or Google Cloud, or only via GitHub and Colab?
- Which operational baselines does the lead-time advantage measure against?