WeatherNext 3 brings hourly 5-km AI forecasts to BigQuery and Maps
DeepMind's sharpest global weather AI is now developer-accessible: hourly forecasts to 15 days, 64-member ensembles, and up to 60% better rain calls.
Copy markdown
Hourly forecasts, five times sharper
WeatherNext 3 drops from WeatherNext 2's 25 km, 6-hour steps to hourly forecasts at 5 km for surface temperature and moisture (10 km for other surface variables, 25 km for winds) — roughly five times the resolution. For builders, that's block-level, hour-by-hour forecasting you can route logistics, energy dispatch, or event planning around.
Rain and temp calls jump up to 60%
Medium-range precipitation improves up to 60% (CRPS) against NASA's IMERG satellite data, ~30% vs radar and ~10% vs rain gauges; 2-meter temperature runs ~40% better than ECMWF at short lead times. That's accurate enough to build real decisions on, not just prettier dashboards.
Query it in BigQuery and Earth Engine now
Access is live: pull forecasts through BigQuery and Earth Engine, bulk-download Zarr files from Google Cloud Storage, or run a custom-inference job on Google Cloud where you set ensemble size and forecast horizon. No physics stack and no supercomputer — the forecast is now a query.
15-day runs, 64-member ensembles
Each forecast is a 15-day ensemble of 64 members from a Functional Generative Network (1024 latent size, 32-layer mesh transformer), trained on data through June 30, 2026. You get probability distributions instead of a single guess — the raw material for risk-weighted planning.
Already live in Search, Maps and Gemini
The same model now powers Google Search, the Gemini app, Maps, and the Google Maps Platform Weather API. If you'd rather not touch Earth Engine, the Maps Platform Weather API is the fastest drop-in for adding weather-aware behavior to an app.