Google WeatherNext 3 is rolling into Search, Gemini and Maps with a fundamental change to AI weather forecasting: it can ingest live satellite observations and generate a fresh global forecast every hour. Google DeepMind and Google Research announced the model on September 3, positioning it as their most advanced global weather system yet.

 

The upgrade pushes key surface variables such as temperature and moisture to 5-kilometer resolution, compared with WeatherNext 2's 25-kilometer grid and six-hour forecast intervals. That combination of finer spatial detail and faster refreshes is designed to improve forecasts for rapidly changing conditions, particularly precipitation.

 

Google WeatherNext 3 Uses Live Satellites

Most AI weather systems have depended heavily on outputs from numerical weather prediction models, which simulate atmospheric physics on supercomputers. Google says those systems can introduce a roughly six-hour data lag before information reaches an AI forecasting pipeline, a meaningful delay for fast-changing rain, clouds and surface temperatures.

 

WeatherNext 3 instead ingests hourly mosaics from geostationary weather satellites alongside historical analysis. Its Functional Generative Network mesh transformer then produces dense global forecast fields, cyclone tracks and station-level predictions from the continuously refreshed observations.

 

Google says the new system adds several operational capabilities:

  • New global forecasts generated every hour
  • Temperature and moisture forecasts at up to 5-kilometer resolution
  • Direct ingestion of live geostationary satellite observations
  • Native cyclone-track and station-level predictions
  • New wind, cloud-cover and solar-radiation variables for clean-energy planning

 

The shift matters because local weather is where global models have traditionally lost detail. Mountain ranges, coastlines and convective storm boundaries can disappear into coarse grids, while stale initialization data can leave a forecast chasing conditions that have already changed.

 

Further Reading

 

Precipitation Forecasts Get Sharper

Rain and snow are among the hardest variables for global forecasting systems because the underlying cloud processes evolve quickly and at scales smaller than many global model grids. Google trained WeatherNext 3 against NASA's IMERG satellite precipitation dataset and its own satellite-radar precipitation reanalysis to improve that weakness.

 

Google reports Continuous Ranked Probability Score improvements of up to 60% against IMERG, 30% against the U.S. Multi-Radar/Multi-Sensor system and 10% against rain-gauge measurements at early lead times. Those figures describe specific evaluation settings rather than a universal accuracy gain across every forecast and location.

 

For forecasts a day or more ahead, Google says people using its products can see precipitation predictions that are up to 50% more accurate, with some of the largest gains expected in regions historically underserved by dense forecasting infrastructure. Independent live evaluation provider Brightband has also ranked WeatherNext 3 among leading operational AI forecasting systems.

 

Search, Gemini and Maps Get WeatherNext

WeatherNext 3 is not staying inside a research demo. Google says the model began powering weather experiences across Google Search, the Gemini app, Google Maps, the Google Maps Platform Weather API and Google Earth Engine on the day of the announcement.

 

Developers, researchers and businesses can also query hourly forecast data through BigQuery and Earth Engine or obtain bulk data from Google Cloud Storage. That gives WeatherNext 3 a distribution advantage that most research forecasting models do not have: its predictions can reach both consumer interfaces and enterprise workflows without users running the model themselves.

 

Renewable-energy operators are a particularly clear target. The model adds variables including wind speeds around turbine heights, high-resolution cloud cover and solar radiation, information that can help operators estimate generation as conditions change throughout the day.

 

AI Weather Forecasting Moves Closer to Observations

The technical direction may prove more consequential than any single accuracy figure. AI weather forecasting has largely learned from the output of traditional numerical systems; WeatherNext 3 pushes the pipeline closer to raw observations by incorporating satellite measurements directly into a high-resolution global model.

 

That does not eliminate conventional meteorology. Historical analyses, observational networks and physics-based forecasting systems remain important parts of the broader weather-data ecosystem, and Google explicitly directs users to national meteorological agencies for official severe-weather warnings and public-safety advisories.

 

Competition is also growing. AI weather startups and research groups are experimenting with their own observation-driven systems, so Google's claims around direct raw-observation forecasting will be tested against rapidly evolving alternatives rather than static numerical baselines.

 

For Google, the immediate advantage is reach. WeatherNext 3 turns a research advance into an hourly data layer that can influence travel planning in Maps, everyday questions in Gemini and Search, developer applications in Cloud, and operational decisions in industries where a few hours of weather uncertainty can carry substantial cost.