Glonce

Google DeepMind's WeatherNext 3 forecasts hourly at up to 5km resolution

modelsThe Glonce editorsGoogle DeepMind🔥76Glonce rating76Impact75Novelty70Relevance85Surprise60Players90Credibility85
Google DeepMind's WeatherNext 3 forecasts hourly at up to 5km resolution

Google DeepMind and Google Research have introduced WeatherNext 3, which they call their most advanced and accurate global weather AI model to date; independent live evaluations from Brightband already rank it as the most accurate global weather model available. Most AI weather models, including Google's own predecessor WeatherNext 2, train mainly on numerical weather prediction (NWP) output, physics-based supercomputer simulations that carry a six-hour data lag and can bias fast-changing variables like rain or surface temperature. WeatherNext 3 instead learns directly from real-time observations: live geostationary satellite mosaics updated every hour, combined with traditional historical analysis and sparse weather-station data, feeding a single Functional Generative Network (FGN) mesh transformer. The result is a continuously updating view of the atmosphere that lets the model generate a fresh global forecast every hour instead of every six.

Resolution varies by variable. WeatherNext 3 resolves key surface variables like temperature and moisture at up to 5-kilometer resolution, other surface variables at 10 kilometers, and atmospheric variables such as wind speed at 25 kilometers; the underlying network also outputs dense gridded fields, discrete cyclone tracks, and natively predicts station-level coordinates. Google describes the overall global picture as roughly five times sharper than WeatherNext 2, which produced forecasts on a 25-kilometer grid updated in 6-hour increments. In a side-by-side comparison of 2-meter temperature over the UK, WeatherNext 3's native 5-kilometer (0.05-degree) output resolves local topography that WeatherNext 2's 25-kilometer (0.25-degree) grid instead renders as a pixelated, over-smoothed picture.

Precipitation gets particular attention, since rain and snow are driven by fast, small-scale cloud processes that AI models have historically forecast as blurry estimates or missed storm boundaries entirely. WeatherNext 3 trains on two high-quality precipitation datasets: NASA's satellite-based IMERG (Integrated Multi-satellite Retrievals for GPM) product and Google's own satellite-radar-based global precipitation reanalysis. In medium-range global forecasts, Google reports Continuous Ranked Probability Score (CRPS) improvements of up to 60% against the IMERG baseline, 30% against MRMS, and 10% against rain-gauge measurements at early lead times. For forecasts a day or more out, Google says people will see up to 50% more accurate precipitation predictions, with the largest gains in regions where forecasts have historically been least reliable. The company points to Latin America, Africa and Asia-Pacific, saying these regions have historically been underserved by high-resolution forecasting due to the immense supercomputing costs of running traditional regional models there, and that the result brings localized, high-fidelity forecasting to billions of people and local businesses in these areas.

WeatherNext 3 also adds forecasts built for renewable-energy planning: 100-meter wind speeds, roughly turbine height, for wind-energy output, plus high-resolution cloud cover and solar radiation levels so solar farms can estimate expected output. Google says this data helps grid operators and renewables developers match clean-energy generation to consumer demand more accurately. Starting today, WeatherNext 3 begins powering weather features in Google Search, the Gemini app, Google Maps, the Google Maps Platform Weather API and Google Earth Engine. Google is also opening the hourly forecast data itself to researchers, developers and businesses, who can query it directly in BigQuery and Google Earth Engine or bulk-download it from Google Cloud Storage, with no model setup required.

Key facts

  • WeatherNext 3 trains on live satellite mosaics, historical analysis and weather-station observations instead of the six-hour-lagged NWP simulations most AI weather models use, generating a new global forecast every hour instead of every six.
  • Resolution varies by variable: key surface variables like temperature and moisture reach up to 5-kilometer resolution, other surface variables 10 kilometers, and atmospheric variables like wind speed 25 kilometers, an overall picture Google calls roughly five times sharper than WeatherNext 2's 25-kilometer grid, updated every 6 hours.
  • Precipitation forecasts improve by up to 60% (CRPS) against the IMERG satellite baseline, 30% against MRMS, and 10% against rain gauges at early lead times; day-or-more-ahead forecasts will be up to 50% more accurate, with the biggest gains in Latin America, Africa and Asia-Pacific.
  • The model also forecasts 100-meter, turbine-height wind speeds plus cloud cover and solar radiation for renewable-energy planning, and natively outputs discrete cyclone tracks and station-level predictions.
  • WeatherNext 3 starts powering Google Search, the Gemini app, Google Maps, the Google Maps Platform Weather API and Google Earth Engine today, with the hourly data also queryable in BigQuery and Google Earth Engine or downloadable in bulk from Google Cloud Storage.

Why it matters

Weather shapes billions of daily decisions, from grabbing an umbrella to running agriculture, supply chains, clean-energy production and national economies, and forecasting has long been limited by how models are trained. Most systems, including Google's own WeatherNext 2, learn from numerical weather prediction (NWP) simulations that lag reality by six hours and struggle with fast-changing variables like rain. WeatherNext 3 instead learns directly from live satellite and weather-station observations and updates every hour instead of every six, closing much of that gap. Independent live evaluations from Brightband already rank it the most accurate global weather model available, and Google says the overall picture is roughly five times sharper than its predecessor, with precipitation forecasts, historically one of the hardest variables to get right, improving by up to 60% on one benchmark.

Who it affects

The update reaches ordinary users automatically: WeatherNext 3 starts powering weather features in Google Search, the Gemini app, Google Maps and the Google Maps Platform Weather API today, and Google says people planning a day or more ahead will see up to 50% more accurate precipitation forecasts. Researchers, developers and businesses can query the hourly data directly in BigQuery and Google Earth Engine, or bulk-download it from Google Cloud Storage. Grid operators and renewables developers get new forecasts for 100-meter, turbine-height wind speeds and solar radiation, useful for matching clean-energy output to demand. Google says the accuracy gains are largest in Latin America, Africa and Asia-Pacific, regions it describes as historically underserved by high-resolution forecasting because of the immense supercomputing costs of running traditional regional models there.

How to use it

There is no separate product or signup: the forecasts arrive inside services people already use, including Search, the Gemini app, Maps and the Maps Platform Weather API, starting today. Developers and businesses can pull the underlying hourly data through BigQuery, Google Earth Engine, or bulk downloads from Google Cloud Storage, with Google saying no model setup is required to integrate it into existing workflows. A companion research paper and a Weather Lab demo let anyone see WeatherNext 3's forecasts visualized in real time, and Brightband's independent leaderboards show where the model currently ranks.

How solid is it

The headline comparisons come from Google's own blog post rather than an independent audit of every figure: the claim that the model's overall picture is roughly five times sharper, and the precipitation gains reported against IMERG, MRMS and rain-gauge baselines, are Google's own numbers. The one third-party element is Brightband, whose independent live evaluations are credited with ranking WeatherNext 3 the most accurate global weather model to date, alongside its own live leaderboards and a Weather Lab demo where the forecasts can be seen in real time. The post links to an accompanying paper for methodology but does not name it, its authors, or a publication venue in the text, and it gives no model size, training-data volume, or compute cost.

Risks and caveats

Google's own disclaimer says the model is not a substitute for official channels: for severe-weather warnings and safety advisories, it still points people to their local meteorological agency or national weather service. The accuracy figures are framed as ceilings, not averages: the CRPS improvements are 'up to' 60%, 30% and 10% against three different baselines, and the day-ahead precipitation gain is 'up to' 50%, so any single forecast may improve by less. Resolution is also uneven rather than uniform: only 'key' surface variables like temperature and moisture reach the finest, 5-kilometer detail, while other surface variables stop at 10 kilometers and atmospheric variables like wind speed at 25 kilometers. The post itself acknowledges that some atmospheric unpredictability will always remain, and it gives no comparison of WeatherNext 3's performance against any non-Google weather model.

“The atmosphere will always retain a degree of unpredictability. However, by training on real-world observations and bypassing traditional modeling constraints, WeatherNext 3 brings us closer to a future where forecasts truly match what is happening on the ground.”

— Google DeepMind and Google Research