Google DeepMind has launched WeatherNext 3, its latest AI model for global weather forecasting, with a major focus on making predictions more detailed, more frequently updated and more useful at a local level.

Announced on September 3, WeatherNext 3 is designed to generate global probabilistic forecasts every hour by combining AI with live observations from geostationary satellites. Google says the model can produce forecasts at resolutions as fine as 5 kilometers for certain surface variables, making it substantially more detailed than its predecessor.

That matters because weather can change dramatically over surprisingly short distances and periods of time. A forecast that looks accurate across a large region can still miss the rain falling over one neighborhood, a temperature difference between a valley and nearby hills, or a storm that develops rapidly during the day.

WeatherNext 3 is Google’s attempt to tackle that problem with a system that continuously incorporates fresh observations rather than relying solely on the processed atmospheric data traditionally used to initialize AI weather models.

WeatherNext 3 updates forecasts every hour

One of the biggest changes in WeatherNext 3 is how frequently it can refresh its forecasts.

Google says the model ingests live global geostationary satellite mosaics and uses them to initialize new forecasts every hour. Its operational system provides 15-day global probabilistic forecasts across 64 ensemble members, while interim hourly runs provide shorter-range forecasts between the main forecast cycles.

That is particularly useful for rapidly developing weather. A storm system, precipitation band or frontal boundary can evolve significantly over several hours, so relying on a forecast generated from older observations can leave it working with an outdated picture of the atmosphere.

WeatherNext 3 instead gets a fresh look at the planet every hour.

The model also produces weather information at multiple resolutions. Google says station-calibrated surface variables such as temperature and humidity can be represented at 5-kilometer resolution, while other surface variables are produced at 10 kilometers and atmospheric variables at 25 kilometers.

For comparison, WeatherNext 2 operated at a 25-kilometer grid and generated forecasts in six-hour increments. Google describes WeatherNext 3 as roughly five times sharper overall.

It is using real observations, not just simulations

The more interesting part of WeatherNext 3 is not simply the higher resolution. It is what the model is learning from.

Earlier AI weather systems have generally relied heavily on analysis data generated by numerical weather prediction systems. Those systems use physics-based simulations running on supercomputers to estimate the state of the atmosphere.

WeatherNext 3 moves closer to the underlying observations by directly ingesting low-latency satellite information. The research paper behind the model describes this as a way to overcome some of the limitations of AI systems that inherit biases from analysis data.

The model combines live satellite mosaics with other datasets, including ECMWF analysis data, historical reanalysis, precipitation observations and weather-station measurements. Google says it is trained directly against station observations as well, allowing it to make localized predictions for variables such as temperature and dew point.

That could be particularly valuable in places where conventional high-resolution regional forecasting is expensive or difficult to operate.

Google specifically points to regions across Latin America, Africa and Asia-Pacific, where the cost of running traditional regional forecasting systems at high resolution can limit availability. WeatherNext 3’s global architecture is intended to provide detailed predictions without requiring a separate regional model for every part of the world.

Rain and snow forecasting get a major upgrade

Precipitation is one of the hardest parts of weather forecasting, and it is an area where WeatherNext 3 is designed to make a noticeable difference.

Google trained the model using precipitation information from NASA’s Integrated Multi-satellite Retrievals for GPM, along with a global precipitation reanalysis based on satellite radar. The company says its evaluations show improvements in precipitation forecasting against several reference datasets.

Google also says that forecasts made a day or more in advance can provide up to 50% more accurate precipitation predictions, with particularly large improvements in regions where forecasting has historically been less reliable.

The underlying research reports that WeatherNext 3 establishes a new state-of-the-art for probabilistic medium-range forecasting skill, while also improving the granularity of the forecasts.

In practical terms, that could mean better predictions for everything from whether rain will arrive at a particular location to how a larger storm system is likely to develop.

Google is putting WeatherNext 3 into its products

This isn’t being presented as a research project that will sit inside Google’s AI lab.

Google says WeatherNext 3 is being integrated into Google Search, the Gemini app, Google Maps, Google Maps Platform, and Google Earth Engine. Developers and organizations can also access operational forecast data through Google Cloud services including BigQuery, Earth Engine and Cloud Storage.

Google is also making the model available for enterprise and developer use, including custom inference on Google Cloud. Developers can control things such as ensemble size and forecast horizons when generating tailored forecasts.

That gives WeatherNext 3 a much wider potential role than simply powering the weather card in Google Search.

For businesses, more accurate and frequently refreshed weather information can feed into decisions around agriculture, transportation, logistics, energy production and emergency response. Renewable energy is another particularly interesting use case because weather variables such as wind and solar conditions directly affect how much electricity can be generated.

Weather forecasting is becoming an AI competition

WeatherNext 3 also highlights how quickly AI weather forecasting is evolving.

Traditional numerical weather prediction remains enormously important and continues to provide the underlying scientific foundation for modern forecasting. But AI models can offer a different approach: instead of repeatedly solving complex atmospheric equations from scratch, they learn patterns from enormous amounts of historical and observational data.

Google’s latest model is increasingly blurring the distinction between those approaches. WeatherNext 3 still uses information from established numerical forecasting systems, but it also incorporates raw satellite observations, station measurements and satellite-derived precipitation directly into its forecasting pipeline.

The result is a system that Google says can deliver forecasts faster while offering significantly greater spatial and temporal detail.

And perhaps the most important change is that the technology is no longer confined to Google’s research environment. WeatherNext 3 is already being rolled into products used by consumers, developers and businesses.

Google has also launched Weather Lab, where users can visualize operational forecasts and track weather systems and cyclone predictions using WeatherNext technology.

For most people, the immediate result may simply be a weather forecast that gets the rain right a little more often. But underneath that familiar forecast is a much bigger shift: Google is betting that AI can make global weather prediction faster, more localized and more responsive to what is actually happening in the atmosphere right now.