Google DeepMind unveils WeatherNext 3, a higher-resolution global weather AI

Google DeepMind on Thursday introduced WeatherNext 3, which it says is its most advanced global weather AI model yet, combining much finer forecast resolution with stronger precipitation forecasting than both its previous system and traditional weather-model baselines. The company also said it is publishing outputs broadly through its own platforms, while placing tighter terms on the newest real-time data.

The launch matters for two concrete reasons readers can evaluate. First, Google says WeatherNext 3 pushes its forecasts from the 25-kilometer scale used in WeatherNext 2 down to as fine as about 5 kilometers for some surface variables, a change that can better capture coastlines, terrain and localized rainfall. Second, at least some of the model’s outputs are already publicly accessible, including through Google Earth Engine, though the freshest real-time feed is not simply unrestricted open data.

According to Google’s updated WeatherNext developer “Research and benchmarks” page, published Sept. 3 UTC, WeatherNext 3 improves on WeatherNext 2 by increasing resolution to 0.05 degrees, or about 5 kilometers, for station-calibrated surface variables and 0.1 degrees, or about 10 kilometers, for gridded surface variables. WeatherNext 2, by comparison, operated at 0.25 degrees, or about 25 kilometers. Google also says the model is initialized every hour and uses live geostationary satellite input.

On performance, Google says WeatherNext 3 delivers “up to a 50% reduction in Brier score and CRPS for precipitation” compared with numerical weather prediction baselines, using global IMERG observations as the evaluation reference. Those are technical measures of forecast quality, and the company is presenting them as evidence that the new system is better at predicting rain. Google’s developer page also points to Brightband’s Operational WeatherBench, a live evaluation service, and says those live evaluations indicate WeatherNext 3 is “the most advanced and accurate global weather model to date.” That assessment is Google’s characterization of Brightband’s results, not an independent confirmation here.

Google linked the release to a technical paper, “WeatherNext 3: Increasing resolution and performance of global weather models with raw observations,” by Remi Rasp and colleagues, dated 2026. That places the announcement within a research program that is producing papers as well as product rollouts, rather than a one-off demo.

The company is also making outputs available in ways that are easy to inspect. A dataset called “WeatherNext 3 (0.05°)” is already listed in the Google Earth Engine catalog, with coverage from Jan. 1, 2026, through at least Sept. 2, 2026, and a pixel size of about 5,566 meters. Google says WeatherNext models are available through several company platforms, but for most users the practical access points are Earth Engine for forecast data and the google-deepmind GitHub repository for code and example notebooks. The repository says the code is published under the Apache 2.0 license.

But the release is not wholly open in the broadest sense, and that is an important part of the announcement. Google’s “GDM Real-Time Weather Forecasting Experimental Data — Terms of Use,” last modified Sept. 3, says: “Any data that relates to a time 1 hour ago or more is licensed under the Creative Commons Attribution International License, Version 4.0 (CC BY 4.0).” Newer real-time experimental data, however, is governed by separate restrictions. In practical terms, that means historical and slightly delayed forecast data can be broadly reused, while the freshest feed comes with tighter conditions.

WeatherNext 3 is the latest step in a longer-running Google forecasting effort that includes GraphCast and WeatherNext 2. It also arrives less than a month after a peer-reviewed Nature paper on WeatherNext Cyclones, published Aug. 6, reported roughly a day or more of lead-time advantage over leading operational models for tropical cyclone forecasting. That paper included authors from Google DeepMind, Google Research, the National Oceanic and Atmospheric Administration’s National Hurricane Center, Colorado State University’s Cooperative Institute for Research in the Atmosphere, and the U.K. Met Office, underscoring that Google’s weather work is increasingly tied to operational forecasting institutions as well as its own platforms.