Google's WeatherNext 3 forecasts every hour at 5 kilometers, and you can query it from BigQuery

Weather models are usually announced as research and reach developers much later. WeatherNext 3, published by Google DeepMind and Google Research on September 3, 2026, arrives the other way round: the forecast data is in BigQuery and Earth Engine on the day of the announcement, and bulk downloads sit in Google Cloud Storage.

What changed in the model

The headline change is what the model reads. Most AI weather models, including WeatherNext 2, train on output from numerical weather prediction systems, and those are physics simulations that carry a six-hour data lag. WeatherNext 3 ingests a mosaic of live global geostationary satellite data instead, so it can emit a new forecast every hour, each one grounded in the most recent observations.

The second change is grain. Key surface variables such as temperature and moisture come out at 5 kilometers, other surface variables at 10, and atmospheric variables such as wind speed at 25. Google describes the overall picture as roughly five times sharper than WeatherNext 2, which forecast on a 25-kilometer grid in six-hour increments. The model also trains directly on sparse weather station observations, which is what lets a global 5-kilometer grid account for a valley or a coastline.

• Hourly forecasts instead of six-hour increments.
• 5-kilometer grid for temperature and moisture, 10 kilometers for other surface variables, 25 kilometers for atmospheric ones.
• 100-meter wind speed, cloud cover, and solar radiation, aimed at wind and solar output estimates.

The precipitation numbers, and who measured them

Rain is where global models have always been weakest, and it is where Google puts its numbers. Against baselines, the Continuous Ranked Probability Score improves by up to 60% against NASA's IMERG, 30% against MRMS, and 10% against rain gauge measurements at early lead times. For people rather than researchers, Google states the day-ahead case separately: up to 50% more accurate precipitation forecasts when planning a day or more in advance.

Those comparisons are Google's own. The broader claim, that this is the most accurate global weather model to date, is attributed to independent live evaluations by Brightband, and that distinction is worth keeping. A vendor's evaluation of its own model and a third party's live leaderboard are two different kinds of evidence, and the post uses both.

Where you can reach it

Access is the part that makes this a tool rather than an announcement. Hourly global predictions are available with no model setup, and the same model starts powering weather in consumer surfaces on the same day.

• Query the forecast data in BigQuery and Google Earth Engine, or bulk-download it from Google Cloud Storage.
• The Google Maps Platform Weather API is served by the new model.
• Google Search, the Gemini app, and Google Maps start using it as of the announcement.