Google has introduced WeatherNext 3, a new global AI weather model that moves some of the most interesting machine-learning progress far away from the chatbot interface.
The system combines historical atmospheric data with live geostationary satellite observations and produces a new forecast every hour. Google says it can generate surface variables at resolutions as fine as 5 kilometres, compared with the 25-kilometre grid and six-hour intervals used by WeatherNext 2.
It is also moving directly into products. Google says WeatherNext 3 is beginning to power weather information in Search, Gemini and Maps, while forecast data is being made available through Google Cloud and Earth Engine.
Why this is technically interesting
Traditional numerical weather prediction relies on large physics-based simulations. These systems divide the atmosphere into a grid and calculate how temperature, pressure, moisture, wind and other variables evolve over time.
AI weather models take a different route. Instead of explicitly solving every physical equation during each forecast, they learn patterns from enormous collections of historical weather data and observations.
WeatherNext 3 pushes this further by ingesting live satellite mosaics rather than relying only on delayed outputs from conventional numerical-weather systems.
That matters because rapidly changing local conditions can be difficult to capture when the freshest input is several hours old.
Higher resolution changes what a forecast can describe
Resolution is more than a prettier map.
Weather conditions can vary sharply around coastlines, valleys, mountains and dense urban areas. A 25-kilometre grid necessarily smooths over many of those differences. Google says WeatherNext 3 can represent selected surface variables at 5-kilometre resolution, other surface variables at 10 kilometres and atmospheric variables at 25 kilometres.
The company also says the model is trained directly on sparse weather-station observations for some variables, helping it learn local effects that broader atmospheric representations can miss.
These are Google’s descriptions of the model and its evaluations, not independent conclusions from Week of Tech. Google points to independent live evaluation by Brightband and provides technical material for deeper inspection.
Precipitation remains the difficult test
Rain and snow are especially challenging because the relevant processes can develop on small spatial and temporal scales.
Google says WeatherNext 3 incorporates satellite-derived precipitation datasets including NASA’s IMERG data and reports substantial improvements on several precipitation metrics compared with its baselines.
Those numbers are useful but should be treated as model-evaluation results rather than a universal guarantee that every local forecast will improve.
The practical test will come from sustained real-world performance across different regions, seasons and weather regimes.
From research model to everyday infrastructure
The most consequential part of the announcement may be deployment rather than architecture.
WeatherNext 3 is being integrated into products used by large numbers of people, while developers and organizations can obtain forecast data through BigQuery, Earth Engine and Cloud Storage.
Google also highlights energy-specific variables such as turbine-height wind speed and solar-radiation estimates. That turns weather prediction into an input for electricity planning, renewable generation and industrial operations rather than simply deciding whether to carry an umbrella.
This is a useful example of where AI may become most important: embedded inside systems people already depend on, often without a conversational interface at all.
What to watch
The key questions are now operational:
- how WeatherNext 3 compares with national meteorological services and leading numerical models over long periods;
- whether improvements hold across regions with sparse ground observations;
- how useful the higher-resolution forecasts prove for energy, agriculture and logistics;
- and how Google communicates uncertainty when AI forecasts enter consumer-facing products.
Google itself cautions users to rely on local meteorological agencies and national weather services for official warnings and public-safety advisories.
That caveat is important. AI weather prediction is becoming infrastructure, but forecasting remains probabilistic and consequential decisions still require authoritative warning systems.