Future TechnologyWeatherNext 3 Explained: Google's AI Forecast Now Updates Every Hour at 5km
Key takeaways
- WeatherNext 3 produces 15 day probabilistic forecasts initialised every hour across 64 ensemble members
- Temperature and moisture arrive at 5km resolution, other surface variables at 10km, atmospheric variables at 25km
- Google claims up to 50 percent more accurate precipitation forecasts a day or more ahead, with the largest gains in poorly observed regions
Sixty four ensemble members, refreshed every hour, running fifteen days ahead. That is the shape of WeatherNext 3, which Google DeepMind and Google Research launched on 3 September. It is already answering the weather question inside Search and Maps.
WeatherNext 3 explained: what the model actually produces
The model outputs 15 day global probabilistic forecasts, initialised hourly. Resolution varies by variable. Temperature and moisture come out at 5km, other surface variables at 10km, and atmospheric variables such as wind at 25km.
The probabilistic part matters. Instead of one deterministic run, WeatherNext 3 generates 64 versions of the next fortnight, which is how you get a usable chance of rain rather than a yes or no that is wrong half the time.
The accuracy claim you can check yourself
Google claims up to 50 percent more accurate precipitation forecasts when you are planning a day or more ahead, with the biggest gains in regions where forecasting has historically been worst.
That last detail is the part worth sitting with. Physics-based forecasting is expensive and depends on dense observation data, so the places with the thinnest weather station networks ended up with the thinnest forecasts. A model that has learned global atmospheric patterns does not care where the supercomputer sits, or how many thermometers happen to be nearby.
It also makes weather an unusual sort of AI claim. Most benchmark results are an argument about methodology. This one gets settled every day by whether it rained.
Where it already runs
WeatherNext 3 is wired into Google Search, the Gemini app, Google Maps and the Maps Platform Weather API. Researchers can pull the data from BigQuery and Earth Engine, or bulk download it from Cloud Storage in Zarr format.
Distribution is the quiet part of this. Gemini passed a billion monthly users earlier this year, so a forecasting model shipped inside Search reaches more people on day one than most national met offices serve in a year.
What to watch
The open question is cost. Physics-based forecasting burns supercomputer time on every run, and AI inference carries its own power bill. Running 64 ensemble members every hour at 5km is not free either. If that trade turns out to favour the model, national forecasting agencies will be having a budget conversation before they have a scientific one.
