What happened?

Google Research and DeepMind have released WeatherNext 3, a weather forecasting model whose core departure from earlier generations is this: it drops traditional physics simulations and learns directly from live satellite data.

Earlier AI weather models, WeatherNext 2 included, were trained on numerical weather prediction (NWP) data. Those simulations run on supercomputers and, according to Google, carry a six-hour delay that introduces errors for fast-changing variables such as rainfall and temperature. WeatherNext 3 instead processes live geostationary satellite data and produces a fresh forecast every hour from the latest observations.

Five times sharper

The model issues hourly forecasts at several resolutions: temperature and humidity at 5 kilometres, other surface variables at 10, and atmospheric values such as wind speed at 25. For comparison, WeatherNext 2 used a 25-kilometre grid at six-hour intervals — roughly five times coarser.

The model also trains on individual weather station data, which lets it capture coastlines, valleys and mountains. According to Google the regions with most to gain are Latin America, Africa and Asia-Pacific, where the high compute cost of traditional regional models has left forecasting underserved.

Rain and renewable energy

Global weather models have long struggled with precipitation, because rain and snow depend on fast cloud processes at small scales. Google trained the model on two sources: NASA's satellite-based IMERG dataset and Google's own global precipitation analysis built on satellite radar.

  • For medium-range global forecasts the CRPS score improves by up to 60 percent over IMERG, 30 percent over MRMS and 10 percent over rain gauges at short lead times.
  • The model predicts wind speed at 100 metres — roughly turbine height — to estimate wind farm output.
  • More accurate cloud cover and solar irradiance values let solar installations calculate expected generation.
  • Users planning a day or more ahead should see precipitation forecasts up to 50 percent more accurate.

Where is it used?

The forecast data updates hourly. Researchers, developers and businesses can query it through BigQuery and Earth Engine or download it in bulk from Google Cloud Storage. WeatherNext 3 now powers the weather features in Google Search, the Gemini app, Google Maps, the Maps Platform Weather API and Earth Engine.

Google draws the limit itself

The company acknowledges that the atmosphere will always remain somewhat unpredictable, and points users to national weather services for official forecasts, severe weather warnings and safety advisories. That distinction matters: the model improves everyday planning without becoming the official warning authority.

To gauge the size of the jump, look at the previous generation. WeatherNext 2, released in November 2025, beat the first generation on 99.9 percent of all meteorological variables and forecast windows while running eight times faster, processing hundreds of weather scenarios in under a minute on a single TPU. In August 2026 DeepMind released WeatherNext 2 and WeatherNext Cyclones, which predicts tropical storm tracks, under an open licence.