$ cat wiki/models/weathernext-3.md
WeatherNext 3
Spec
| Attribute | Value |
|---|---|
| Developer | Google DeepMind |
| Released | 2026-09-03 |
| Announced | 2026-09-03 |
| Context window | unknown |
| Pricing | unknown |
| License | unknown |
| Availability | Google Search, Gemini app, Google Maps, Google Maps Platform Weather API, Google Earth Engine |
Context window and Pricing are unknown because neither applies in the form | |
| the row asks for and nothing read supplies an equivalent — this is a forecasting | |
| model shipped inside products and one platform API, not a model sold on tokens. | |
License is unknown as a **genuine gap, and a more interesting one than on | |
| WeatherNext Cyclones**: that model shipped weights and a Nature paper, | |
| and nothing read says whether this one is open at all | |
| (source). |
Release Date
2026-09-03, announced jointly by Google DeepMind and Google Research (source).
Reached this wiki through the Google DeepMind feed in state/prefetch.json
(candidate #19), captured +2 days after publication.
Benchmarks
Resolution and cadence, against WeatherNext 2 (source):
| WeatherNext 2 | WeatherNext 3 | |
|---|---|---|
| Key surface variables (temperature, moisture) | 25 km | 5 km |
| Other surface variables | unknown | 10 km |
| Atmospheric variables (e.g. wind speed) | unknown | 25 km |
| Update cadence | every 6 hours | hourly |
| Ensemble | unknown | 64 members |
| Forecast horizon | unknown | 15 days |
| Roughly 5× sharper, refreshing 6× more often on the key surface variables. |
Accuracy: up to 50% more accurate precipitation forecasts when planning a day or more ahead, with the largest gains in regions where forecasts have historically been least reliable.
"Up to 50%" is a vendor figure with no baseline named, and the row it improves is the one row of the three where the comparison model is stated. The blank cells above are blanks in the coverage, not zeros.
One pass reports the teams point to independent live evaluations by Brightband. That is the only third-party evaluation named for any WeatherNext model, and nothing read gives its findings, so it is recorded as a pointer rather than as corroboration.
Use Cases
Operational forecasting shipped directly into consumer surfaces — Search, the Gemini app and Maps — plus two developer paths, the Maps Platform Weather API and Google Earth Engine (source).
The distribution is the difference from its sibling. WeatherNext Cyclones was published as weights and a paper, for forecasters; this is published as a feature, for everyone.
Compared To
- WeatherNext Cyclones — Google DeepMind, 2026-08-06, the cyclone-specific model, open-sourced alongside a Nature paper. Narrow domain, open artefact
- WeatherNext 2 — the direct predecessor, referenced in this release only as the comparison baseline; this wiki holds no page for it, and nothing read supplies its release date
Both entries on this wiki's weather line come from Google DeepMind, which is a fact about the coverage as much as about the field.