September 3, 2026
Our flagship AI weather forecasting model now incorporates real-time satellite data, hourly updates, higher resolution, precise precipitation forecasts, and clean energy variables. The model is now integrated into Google Search, Gemini, Maps, Google Maps Platform, and Cloud.
The WeatherNext Team
Weather affects the decisions of billions of people every day. Some decisions are simple, such as taking an umbrella before heading out; others have far-reaching consequences. Wind, rainfall, and extreme weather events such as heat waves and droughts can trigger cascading effects across agriculture, global supply chains, clean energy production, and national economies.
In recent years, AI has transformed weather forecasting, using historical records to make predictions faster and more accurately than traditional methods. However, forecasting highly localized and rapidly changing weather remains a challenge. Earlier models typically lacked sufficient spatial resolution and struggled to incorporate real-time weather data from sources such as satellites.
Today, Google DeepMind and Google Research are introducing WeatherNext 3. According to Brightband’s independent live evaluation, it is the most advanced and accurate global weather model to date. The model learns directly from real-time observations, enabling timely, more localized forecasts of the weather events that affect people most. By using raw satellite data to generate high-resolution forecasts every hour, the model brings reliable weather forecasting to Google products around the world.
Fast Weather Forecasting at Unprecedented Resolution
The usefulness of a weather forecast often depends on the level of detail it provides and its ability to resolve time and space at a fine scale. WeatherNext 3 can generate hourly forecasts at multiple spatial resolutions while maintaining physical consistency between broad global wind patterns and local terrain.
With WeatherNext 3, we can visualize key surface variables such as temperature and humidity at a resolution of 5 kilometers, represent other surface variables at 10 kilometers, and show atmospheric variables such as wind speed at 25 kilometers. Overall, this provides a view of global weather that is approximately five times sharper than that of our previous-generation model, WeatherNext 2. WeatherNext 2 generated forecasts on a 25-kilometer grid at six-hour intervals.
Figure 1: End-to-end system architecture for WeatherNext 3. The model combines real-time, one-hour geostationary weather satellite mosaics with traditional historical analysis data and feeds them into a flexible Functional Generative Network (FGN) grid Transformer. The model outputs dense gridded fields, discrete cyclone tracks, and natively forecasts sparse station-level coordinates.
Figure 2: Comparison of 2-meter air temperature forecasts over the United Kingdom. The left panel shows WeatherNext 2 at a resolution of 25 kilometers (0.25°); the right panel shows WeatherNext 3 at its native resolution of 5 kilometers (0.05°). WeatherNext 3 can resolve complex local terrain, avoiding the pixelation and overly smoothed temperature patterns seen in older models.
Continuous Global Coverage with Constantly Updated Real-World Data
The biggest advance in WeatherNext 3 lies in the data it learns from. Most AI weather models, including WeatherNext 2, are trained on data generated by numerical weather prediction (NWP) models. While NWP models are highly useful, they are complex physical simulations powered by supercomputers and have a six-hour data latency. For rapidly changing variables such as precipitation and surface temperature, this delay can introduce biases.
By ingesting real-time global geostationary weather satellite mosaics, our new model receives a rich, continuously updated view of the atmosphere. This allows the model to generate a new forecast every hour, with each forecast based on the latest available satellite observations at resolutions of up to 5 kilometers.
This is particularly important because critical weather conditions can develop quickly. When storms, fronts, or precipitation systems suddenly form, a faster update cycle and higher resolution can provide more detailed information sooner, helping drive an effective response.
Variables such as temperature and humidity can change dramatically over just a few kilometers, which is especially important for communities near coastlines, valleys, or mountain ranges. Traditional models perform poorly in these situations because they are trained on atmospheric representations that lack detail, making them prone to missing localized extreme variations.
To address this challenge, WeatherNext 3 is trained instead using sparse observations from weather stations. This enables us to generate global forecasts on a 5-kilometer grid while incorporating regional details such as terrain.
This breakthrough is particularly important for many areas across Latin America, Africa, and the Asia-Pacific region. These areas have historically lacked access to high-resolution weather forecasts because traditional regional models require costly supercomputing resources. WeatherNext 3 brings localized, high-fidelity weather forecasts to billions of people and local businesses in these regions.
In addition to improving resolution and forecast frequency, our model is specifically equipped with forecasting capabilities for renewable energy production. It can predict wind speeds at an altitude of 100 meters—approximately the height of a wind turbine—enabling precise estimates of wind power generation. It also provides high-resolution cloud cover and solar irradiance levels, helping solar farms estimate how much sunlight will reach the ground.
This data is critical for clean energy planning worldwide. Grid operators and renewable energy developers can use it to accurately forecast power generation from clean energy assets and match it with consumer demand.
Breakthroughs in Precipitation Forecasting Accuracy
Global weather models have long struggled to accurately predict precipitation. Rain and snow systems are driven by rapidly moving, small-scale cloud processes that are difficult for traditional physics-based simulations to model accurately. As a result, AI forecasts often produce blurry estimates or fail to identify the boundaries of intense storms altogether.
To address this challenge, we trained the model on two exceptionally high-quality precipitation datasets: NASA’s Integrated Multi-satellite Retrievals for GPM (IMERG), a satellite-based global precipitation measurement product, and our global precipitation reanalysis generated from satellite radar.
The results show a substantial improvement in precipitation forecast accuracy. For medium-range global forecasts, evaluations show that the Continuous Ranked Probability Score (CRPS) improves by up to 60% relative to IMERG, 30% relative to MRMS, and 10% relative to rain-gauge measurements at shorter lead times.
Figure 3: Comparison of medium-range precipitation probability (PoP > 1mm) forecasts. The left panel shows WeatherNext 2 at a resolution of 25 kilometers, with a highly diffuse and pixelated precipitation area; the center panel shows WeatherNext 3 at a resolution of 11 kilometers, closely matching the actual satellite-based ground truth (right panel) and accurately capturing t