The WeatherNext 3 Model, Explained
WeatherNext 3 is Google DeepMind's third-generation weather model: an hourly, higher-resolution system that reads raw satellite observations directly and feeds forecasts into Google's own products. This page covers what changed, what it produces and how to judge the claims. For the practical side, see how to read the forecast.
Independent third-party analysis. WeatherNext and Google DeepMind are trademarks of their respective owners.
WeatherNext 3 by the numbers
Three figures separate this generation from the one before it.
1 h Global forecasts are generated hourly rather than on a six-hourly cycle, so fast-changing conditions appear sooner.
Global forecasts are generated hourly rather than on a six-hourly cycle, so fast-changing conditions appear sooner.
5 km Temperature and humidity resolve at 5 km, other surface variables at 10 km and upper-air levels around 25 km.
Temperature and humidity resolve at 5 km, other surface variables at 10 km and upper-air levels around 25 km.
60% Reported improvement in precipitation skill against the IMERG reference in Google's own evaluation.
Reported improvement in precipitation skill against the IMERG reference in Google's own evaluation.
What actually changed from WeatherNext 2
WeatherNext 2 ran at roughly 25 km with six-hourly updates. The jump to hourly output at 5 km for the sharpest variables changes which decisions the forecast can support, and it comes mainly from how the model takes in observations.
Raw satellite ingestion
The model consumes geostationary satellite observations directly instead of relying only on pre-processed analyses.
Hourly generation
New global output every hour keeps a live operation on the newest run rather than the newest daily cycle.
Sharper local detail
Targeted temperature and humidity at 5 km, with most other surface variables at 10 km.
Energy variables
Wind at hub height, cloud cover and radiation are produced with renewable planning in mind.
How to judge the claims
A model announcement is not the same as an independent evaluation. These are the things worth checking before you rely on a number.
Read the paper
The technical write-up documents the method, the training data and the evaluation protocol behind the headline figures.
Check the baseline
Skill numbers are always relative to something. Know whether the comparison is the previous generation, a physical model or a satellite product.
Separate skill from speed
Faster generation matters operationally, but it is a different claim from better accuracy.
Look at regional behaviour
Global averages hide weak spots. Performance in data-sparse regions and around extremes deserves its own check.
Keep the limits in view
No AI forecast replaces official warnings. Treat model output as the planning layer inside those boundaries.
Test on your own decisions
The only benchmark that pays is your own outcome: did the hourly forecast change a decision for the better?
WeatherNext 3 model questions
The technical questions that come up most often.
Use the model without running it yourself
Hourly forecasts, local variables and one workspace, starting with a 7-day free trial. See the step-by-step guide to get productive on the first attempt.