Google DeepMind and Google Research released WeatherNext 3 today, an AI weather model that generates global forecasts every hour at 5-kilometer resolution for surface variables like temperature and moisture. That is five times sharper than WeatherNext 2, which produced forecasts on a 25-kilometer grid in 6-hour increments. The model is rolling out immediately across Google Search, the Gemini app, Google Maps, the Google Maps Platform Weather API, and Google Earth Engine.
The technical leap is what the model learns from. WeatherNext 2 and most other AI weather systems train on outputs from numerical weather prediction models — physics simulations that require supercomputers and carry a six-hour lag. WeatherNext 3 instead ingests live geostationary satellite mosaics at 1-hour cadence, producing a new forecast every hour grounded in the most recent observations.
“Our model learns directly from real-time observations, enabling it to provide timely and more localized predictions for the weather events that impact people the most.”— The WeatherNext team, Google DeepMind and Google Research
Resolution scales by variable. Key surface fields resolve at 5 kilometers (0.05° native), other surface variables at 10 kilometers, and atmospheric variables like wind speed at 25 kilometers. Independent live evaluations by Brightband rank WeatherNext 3 as the most accurate global weather model to date, according to Google's announcement.
Key facts
- 01WeatherNext 3 produces hourly forecasts at 5-kilometer resolution for temperature and moisture, five times sharper than WeatherNext 2's 25-kilometer grid.
- 02The model trains on live geostationary satellite mosaics instead of numerical weather prediction outputs, eliminating the six-hour data lag.
- 03Precipitation forecasting improves by up to 60% on CRPS against NASA's IMERG benchmark and 30% against MRMS.
- 04Longer-term forecasts a day or more out gain up to 50% in accuracy over the prior model.
- 05The system rolls out today across [Google](/gemini) Search, Gemini, Maps, Maps Platform Weather API, Earth Engine, and Google Cloud.
The architecture centers on a Functional Generative Network mesh transformer that outputs dense gridded fields, discrete cyclone tracks, and station-level predictions natively. It trains directly on sparse weather station observations, which lets it capture regional detail — coastlines, valleys, mountain ranges — that low-resolution atmospheric representations flatten out.
Precipitation is where the numbers land hardest. Global models have long struggled with rain and snow because the driving cloud processes operate at scales physics simulations cannot resolve cleanly. Trained on NASA's Integrated Multi-satellite Retrievals for GPM (IMERG) and Google's own global precipitation reanalysis, WeatherNext 3 posts a Continuous Ranked Probability Score improvement of up to 60% against IMERG, 30% against MRMS, and 10% against rain gauge measurements at early lead times. Medium-range probability-of-precipitation forecasts render at 11-kilometer resolution.
The model also introduces variables built specifically for renewable energy operators. It forecasts 100-meter wind speeds at roughly turbine height, alongside high-resolution cloud cover and solar radiation figures for photovoltaic output planning. Grid operators can use those inputs to match generation against consumer demand — a persistent scheduling problem as clean energy assets grow as a share of supply.
“This breakthrough is particularly vital for regions across Latin America, Africa, and Asia-Pacific that have historically been underserved by high-resolution forecasting due to the immense supercomputing costs of traditional regional models.”— The WeatherNext team, Google DeepMind and Google Research
Longer-range forecasts improve as much as short-range ones. Google says users planning a day or more ahead will see up to 50% more accuracy compared to the prior model. That gain matters for logistics, agriculture, and emergency planning, where a decision window of 24 to 72 hours is standard.
Distribution is aggressive. Beyond consumer surfaces in Search, Maps, and the Gemini app, developers and researchers can query WeatherNext 3 data directly in BigQuery and Earth Engine, or bulk-download from Google Cloud Storage. No model setup is required — Google is treating the forecasts as a data product, not just a feature.
The independent-evaluation claim rests on Brightband's live scoring, and rival forecasting shops from the European Centre for Medium-Range Weather Forecasts to Nvidia's own weather models will publish their own comparisons. Satellite-driven models also inherit the failure modes of their inputs — outages, calibration drift, and coverage gaps in polar regions can degrade quality in ways that physics-based ensembles handle more gracefully. And precipitation gains against IMERG do not automatically translate to gains against ground-truth radar in every region.
The commercial subtext is the interesting part. Google is positioning weather intelligence as an ambient capability inside its consumer and cloud products, which reframes what has historically been a public-sector data monopoly. If a Maps Platform API call returns a 5-kilometer hourly forecast anywhere on earth for renewable-energy planners, air traffic controllers, and insurers, national meteorological agencies suddenly compete with a hyperscaler on both accuracy and distribution. That is a structural shift in a market Google was not previously seen as owning.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
