WeatherNext 3, Google DeepMind's Latest Forecasting Model, Sharpens Global Weather Into a 5-Kilometer Picture

Google DeepMind has released WeatherNext 3, a forecasting model that ingests live global geostationary satellite data and produces a global weather picture roughly five times sharper than its predecessor, WeatherNext 2. The jump matters less as a benchmark score and more as a shift in what kind of decisions weather data can now support.

Three resolutions, one hourly cycle

WeatherNext 3 doesn't output a single flat grid. Temperature and moisture arrive at 5-kilometer resolution, other surface variables at 10 kilometers, and atmospheric variables at 25 kilometers, and the whole set refreshes every hour. That tiered structure means the model is producing several different views of the atmosphere simultaneously rather than one uniform snapshot, and it's doing so on a cadence tight enough to track how conditions shift through a single day rather than just across days. The model was trained on NASA's IMERG precipitation data and directly on sparse weather-station observations, which is part of how DeepMind gets a globally consistent picture even in regions where ground instrumentation is thin.

Where the accuracy gains actually show up

DeepMind reports CRPS improvements of up to 60% against IMERG and 30% against MRMS in medium-range global forecasts, with a 10% improvement against rain-gauge measurements specifically at early lead times. Those are three different reference datasets measuring three different aspects of precipitation forecasting, and the fact that the gains hold across all three is what makes the number credible rather than cherry-picked. On top of that, DeepMind states that planning one or more days ahead now benefits from precipitation forecasts up to 50% more accurate. That's the kind of improvement that changes what a forecast can be used for, not just how good it looks on a leaderboard.

Built for energy grids as much as for umbrellas

Two of WeatherNext 3's outputs are aimed squarely at energy operations: wind speed forecasts at 100 meters, the height that matters for turbine output, and cloud cover and solar radiation forecasts for solar generation planning. Pairing those with hourly refresh and satellite ingestion means grid operators get a forecasting layer built around the variables that actually drive generation forecasts, rather than a general-purpose weather model that happens to include wind and cloud fields as an afterthought.

Getting the data out of DeepMind's hands and into yours

WeatherNext 3 is already surfaced inside consumer-facing Google products, including Google Search, Gemini, Google Maps, the Google Maps Platform Weather API, and Google Earth Engine. For anyone who needs the raw forecasts rather than a rendered map, the data can be queried through BigQuery and Earth Engine, or bulk-downloaded through Google Cloud Storage. That gives three distinct access patterns: interactive querying against BigQuery for analysts who want to join weather fields with other datasets, geospatial analysis through Earth Engine for anyone already working in that environment, and direct bulk retrieval from Cloud Storage for teams building their own pipelines. Which of those three fits depends entirely on what a given team is already using, which is the real question worth answering before touching the data at all.

Specs

• Forecast refresh interval: Hourly
• Temperature and moisture resolution: 5 kilometers
• Other surface-variable resolution: 10 kilometers
• Atmospheric-variable resolution: 25 kilometers
• Wind forecast height: 100 meters
• Precipitation CRPS improvement vs IMERG: Up to 60%
• Precipitation CRPS improvement vs MRMS: 30%
• Precipitation CRPS improvement vs rain gauges: 10%

Sources

• WeatherNext 3 — Google DeepMind