Built for meteorologists, renewable-energy forecasters, and catastrophe modelers who need sharper, faster output than legacy numerical weather prediction, WeatherNext 3 scales to roughly 2.4 times WeatherNext 2's parameters and runs a 64-member ensemble out to a 15-day horizon. It is not a substitute for official severe-weather warnings from a national meteorological agency.
Launched September 3, 2026, WeatherNext 3 is the global AI weather model Google DeepMind builds a new forecast from every hour using live satellite data, instead of waiting on six-hourly physics simulations. It produces station-level, cyclone-track, and gridded atmospheric forecasts in a single ensemble run, sharper and faster than the WeatherNext generation it replaces.
Provider: Google DeepMind · Family: WeatherNext
More about Google DeepMind on HokAI
Input modalities: satellite-imagery, weather-station-observations, atmospheric-reanalysis-data, numerical-weather-fields · Output: gridded-forecast-fields, station-point-forecasts, tropical-cyclone-tracks
About WeatherNext 3
WeatherNext 3 is Google DeepMind's global AI weather forecasting model, built with Google Research and launched on September 3, 2026. It uses a Functional Generative Network (FGN) mesh transformer, an encode-process-decode architecture that maps separate input encoders for each data source onto a shared icosahedral processor mesh, with a latent dimension scaled from 768 to 1024 and a processor depth scaled from 24 to 32 transformer layers versus WeatherNext 2. Where physics-based numerical weather prediction needs a supercomputer and only updates every six hours, WeatherNext 3 is designed to run inference on a single TPU and refresh far more often, solving the core problem of stale, coarse forecasts between synoptic cycles. Google DeepMind reports up to a 60 percent reduction in precipitation CRPS (Continuous Ranked Probability Score, the standard measure of probabilistic forecast error) against IMERG satellite ground truth at early lead times, roughly 30 percent against radar-based MRMS estimates, and up to a 50 percent reduction in Brier score for extreme-precipitation events, all measured against WeatherNext 2. Independent evaluator Brightband, which runs the Operational WeatherBench benchmark, rated WeatherNext 3 its most accurate global weather model this cycle, placing it above rival AI systems and the traditional forecasts national weather services still rely on. The model runs on three nested resolutions. Station-point temperature and dew point are trained directly on raw weather-station measurements at 0.05 degrees, roughly 5 kilometers, a sharper grid than any prior WeatherNext release. Gridded surface fields, including wind, pressure, sea-surface temperature, cloud cover, and precipitation, resolve at 0.1 degrees (about 10 kilometers), while atmospheric fields across 13 pressure levels resolve at 0.25 degrees (about 25 kilometers). Its main synoptic cycles, initialized four times daily, forecast out to 15 days; interim hourly cycles covering the other twenty hours of the day forecast out to 48 hours, so a fresh run launches every hour rather than every six. WeatherNext 3 trains on ERA5 and HRES-fc0 atmospheric reanalysis, NASA's IMERG satellite precipitation retrievals, Google's own satellite-radar precipitation reanalysis, and raw weather-station observations, rather than only gridded reanalysis products. At inference time it additionally ingests live, one-hour-old geostationary satellite mosaics alongside ECMWF HRES analysis fields, which is what lets it assimilate real-world conditions instead of waiting on the next physics-model run. Output covers dense gridded fields, station-level point forecasts, and discrete tropical-cyclone tracks (formation, track, intensity, size, and shape) through the related WeatherNext Cyclones component. There is no per-token or per-forecast API price. Historical forecast output, anything one hour old or older, is licensed under Creative Commons Attribution 4.0 and is free to query from BigQuery, Earth Engine, and Cloud Storage, though normal Google Cloud storage and query costs still apply to bulk downloads. Real-time and future forecasts fall under the separate GDM Real-Time Weather Forecasting Experimental Data Terms of Use, which brands the product experimental. Consumers get WeatherNext 3 forecasts for free inside Google Search, the Gemini app, and Google Maps, with no separate subscription. Custom, on-demand inference outside the published forecast cadence is gated behind a Google Cloud allowlist, and unapproved requests currently fall back to serving WeatherNext 2 output instead of WeatherNext 3. WeatherNext 2 itself, along with the older GraphCast and GenCast research models, is fully open on GitHub under Apache 2.0 for the code and CC BY 4.0 for the pretrained weights and other assets; WeatherNext 3's own weights had not been open-sourced as of its September 2026 launch. WeatherNext 3 is not a language model, so it carries no refusal behavior or content-moderation controls; its safety statement is meteorological rather than conversational. Google DeepMind researcher Ferran Alet has said the atmosphere "will always retain a degree of unpredictability," and the product's own terms state it is not a substitute for warnings from a national meteorological agency. The WeatherNext 3 research paper, submitted by 25 Google DeepMind and Google Research authors led by Stephan Rasp, also discloses a concrete limitation: individual ensemble members show subtle hexagonal artifacts and discontinuities at 6-hour forecast boundaries, an effect of the icosahedral mesh, though the authors report it largely disappears once statistics are aggregated across the full ensemble. It is best suited to meteorologists and climate researchers benchmarking global models, renewable-energy forecasters who need wind and solar irradiance data, catastrophe modelers tracking cyclone risk, and developers already building on Google Cloud who can query BigQuery, Earth Engine, or Cloud Storage directly. It is the wrong tool for anyone who needs an official, legally-recognized severe-weather warning, since the terms of use explicitly disclaim that use, and for anyone who needs open, self-hostable weights today, since only the WeatherNext 2 generation is currently public. WeatherNext 3 succeeds WeatherNext 2, announced November 17, 2025, which itself replaced the original WeatherNext package built from GraphCast (a deterministic graph neural network published in November 2023) and GenCast (a probabilistic diffusion ensemble published in December 2024). WeatherNext 2 ran about eight times faster than that prior generation and beat it on 99.9 percent of measured variables and lead times; WeatherNext 3 keeps the same FGN family of architecture but scales it up, and, per Google's own comparison reported by TechCrunch, carries roughly 2.4 times WeatherNext 2's parameter count.
Pricing
No per-token or per-forecast API price is published. Older query results carry an open Creative Commons license and cost nothing beyond ordinary BigQuery, Earth Engine, and Cloud Storage usage fees. Fresh and future output falls under a separate experimental terms-of-use grant, and custom inference beyond the published cadence needs Google Cloud allowlist approval first.
Key Features
- Hourly Global Initialization: Launches a new forecast every hour, 24 times a day, instead of the six-hourly cadence used by physics-based models and WeatherNext 2.
- 5km Station-Level Resolution: Trains directly on raw weather-station temperature and dew point observations at 0.05 degrees, about 5 kilometers, its sharpest global output grid yet.
- 64-Member Probabilistic Ensemble: Produces 64 ensemble scenarios per run, covering gridded fields, 13 atmospheric pressure levels, and discrete tropical-cyclone tracks in one pass.
- Live Satellite Data Assimilation: Ingests live, one-hour geostationary satellite mosaics at inference time alongside ECMWF HRES analysis, rather than relying only on six-hourly reanalysis.
- Renewable-Energy Output Variables: Outputs turbine-height wind speed, solar irradiance, and cloud-layer cover aimed at wind and solar generation forecasting.
Pros
- Outputs turbine-height (100-meter) wind speed and solar irradiance data purpose-built for renewable-energy generation forecasting.
- Consumer access is free inside Google Search, Maps, and the Gemini app, with no separate WeatherNext subscription.
- Trains directly on raw weather-station observations rather than only reanalysis grids, a design difference from most global AI weather models.
- Produces discrete tropical-cyclone tracks, intensity, and size estimates alongside its gridded forecasts, rather than surface variables alone.
Cons
- WeatherNext 3's own model weights are not open-sourced; custom inference currently defaults back to WeatherNext 2 until Google approves allowlist access.
- No published parameter count, training-data cutoff, or governance/system-card documentation exists for WeatherNext 3, unlike Google's LLM releases.
- Real-time forecasts run under an experimental terms-of-use license and are explicitly not a substitute for official warnings from a national meteorological agency.
Frequently Asked Questions
How much does WeatherNext 3 cost to access in 2026?
There is no per-token or per-forecast price list. Anything Google DeepMind considers current output, one hour old or newer, sits under a separate experimental terms grant rather than a paid tier, while everything older is public under an open Creative Commons license, so the only real cost is ordinary Google Cloud usage on BigQuery, Earth Engine, or Cloud Storage. Viewing forecasts inside Search, Maps, or the Gemini app costs nothing at all.
How does WeatherNext 3 compare to WeatherNext 2 on accuracy?
WeatherNext 3 cuts precipitation forecast error (CRPS) by up to 60% versus WeatherNext 2 against satellite ground truth at early lead times, and reduces station-level temperature error by roughly 30%, per Google DeepMind's own published figures. Independent evaluator Brightband separately rated it the strongest global model in its Operational WeatherBench tests as of September 2026.
Is WeatherNext 3 open source?
No. WeatherNext 3's model weights are proprietary and were not publicly released at its September 2026 launch, unlike WeatherNext 2 and the older GraphCast and GenCast models, whose code and weights are open on GitHub under Apache 2.0 and a Creative Commons Attribution 4.0 license. The forecast output itself is more open than the weights: older records carry that same Creative Commons terms, while anything current sits under a separate experimental grant.
Does WeatherNext 3 train on user data?
No. Its inputs are public meteorological sources, reanalysis archives, satellite retrievals, and station instrument readings, never prompts, queries, or files submitted by a developer or consumer. Google has not published a training cutoff date or a dedicated governance write-up for this particular model.
Who is WeatherNext 3 best for, and who should avoid it?
Meteorologists, energy planners tracking turbine and solar output, and catastrophe modelers following storm risk get the most out of it. Skip it if you need a legally recognized severe-weather alert, since the terms explicitly rule that use out, or if you need to run inference on your own hardware today, since only the prior WeatherNext 2 generation has public weights.