WeatherNext 3, the third global weather model from Google DeepMind and Google Research, now learns from live satellite images and weather-station readings instead of only from the six-hour-old output of physics-based forecasts. On the public leaderboard that scores it against rivals every day, it also ranks first overall, ahead of its own predecessor, WeatherNext 2, by 0.02 K of error.
What changed: raw observations instead of another model’s output
Google DeepMind and Google Research announced WeatherNext 3 on 3 September 2026, with a companion paper on arXiv the same day. The model is a single Functional Generative Network, a mesh transformer fed hourly by geostationary satellite mosaics and historical analysis data, and it outputs gridded forecasts, cyclone tracks and station-level predictions.
That input is the core change. WeatherNext 2 trained on the six-hourly output of physics-based analysis models. WeatherNext 3 ingests low-latency satellite data directly, which lets it issue a new global forecast every hour rather than every six. Resolution for surface temperature and moisture reaches 5 km, versus WeatherNext 2’s 25 km grid updated every six hours — Google describes the jump as “roughly five times sharper.” The paper reports 0.1-degree resolution with hourly time steps for single-level variables including solar radiation and cloud cover, which it calls on par with physics-based global models.
Precipitation forecasts are trained on NASA’s IMERG satellite dataset and on Google’s own satellite-radar precipitation reanalysis. The model also adds two variables aimed at energy forecasting: turbine-height (100 m) wind speed and surface solar radiation. Station data additionally lets it predict 2-meter temperature and dewpoint at specific locations, including stations it was not trained on, according to the paper — an authors’ own evaluation not yet independently replicated.
Where it runs, and the numbers Google is not showing on the leaderboard
WeatherNext 3 is already live in Google Search, the Gemini app, Google Maps, the Maps Platform Weather API and Earth Engine, and its output can be queried in BigQuery and Earth Engine or downloaded from Cloud Storage. Google says the model brings “up to 50% more accurate precipitation forecasts” a day or more ahead, with the largest gains in regions it names as historically underserved by high-resolution forecasting: Latin America, Africa and Asia-Pacific. The company also reports CRPS improvements for medium-range precipitation of up to 60% against IMERG, 30% against the MRMS radar product and 10% against rain gauges “for early lead times,” without naming the baseline model in that sentence. Both figures are Google’s own, and neither appears broken out on the public leaderboard described below.
The “most accurate” claim, checked against a public scoreboard
Google calls WeatherNext 3 the most accurate global model “according to independent live evaluations by Brightband.” Brightband’s Operational WeatherBench does score AI and physics models daily against what the atmosphere actually does, using Google Research’s own open WeatherBench-X toolkit — a detail worth reading “independent” against: third-party scoring, on a Google-authored standard.
As consulted on 12 September 2026, with the latest ranked cycle from 4 September at 12Z (the board trails real time by about a week while it waits for verifying data), WeatherNext 3 does rank first on the leaderboard’s headline metric, 850 hPa temperature RMSE over a seven-day forecast: 1.31 K, winning 17 of the last 29 daily cycles. WeatherNext 2 sits second at 1.33 K with 11 wins. ECMWF’s AI ensemble, AIFS-ENS, follows at 1.40 K, then its physics ensemble IFS-ENS at 1.52 K, its high-resolution HRES run at 1.80 K, and NOAA’s GFS at 2.11 K. By region, WeatherNext 3 wins the tropics, both hemispheres and North America. WeatherNext 2 still wins Europe, though, with 9 of 25 cycles on the same metric.
Two claims, checked separately
Weather forecasting is scored in public every day, so the two claims in this launch can be checked separately. The architectural one holds up in the paper: an AI model trained on raw satellites and stations rather than on another model’s six-hour-old analysis, updating hourly at resolutions comparable to physics-based global models. The “most accurate” one holds up on the leaderboard’s own headline number, but the margin over WeatherNext 2 is 0.02 K, the predecessor still wins in Europe, and the precipitation and local-temperature gains Google highlights are not the metric the leaderboard shows.
The broader ranking is the more durable signal: the top four models on Brightband’s board are all AI systems, with ECMWF’s physics ensemble fifth and its flagship HRES run ninth. That would have been the headline two years ago; now it is the baseline against which WeatherNext 3 measured a 0.02 K lead. What that average does not cover — skill on cyclones, on weather extremes the models have rarely seen, and the compute cost of running them — is not published in the sources reviewed for this launch.