Google DeepMind Unveils WeatherNext 3 for Higher-Resolution Global Forecasts

Weather forecasts shape decisions far beyond whether someone should carry an umbrella. Agriculture, power-grid operations, supply chains and emergency response all depend on estimates of wind, rain, temperature and extreme events. Google DeepMind and Google Research have introduced WeatherNext 3, a new global AI weather model aimed at improving both the speed and local detail of forecasting. Google describes it as its most advanced and accurate global weather model so far, citing independent live evaluations by Brightband.

Key points

• Hourly forecasting: WeatherNext 3 ingests mosaics of live observations from geostationary satellites. This gives the system a more continuously updated view of the atmosphere and enables a new forecast every hour.
• Finer grids: Temperature and moisture can be visualized at a 5-kilometer resolution, other surface variables at 10 kilometers, and atmospheric variables such as wind speed at 25 kilometers. Google says the overall picture is roughly five times sharper than WeatherNext 2, which used a 25-kilometer grid and six-hour increments.
• More local context: The model is also trained directly on sparse weather-station observations. The goal is to capture regional effects from coastlines, valleys and mountain terrain that can be smoothed out in coarser global models.
• A focus on precipitation: Rain and snow remain difficult to forecast because they depend on rapidly evolving cloud processes at small scales. WeatherNext 3 is designed to improve precipitation predictions and better preserve the boundaries of intense weather systems.
• Renewable-energy variables: The system forecasts wind at roughly 100 meters above the surface, a height relevant to wind turbines. It also predicts cloud cover and solar radiation to help solar operators estimate available sunlight.

Why it matters

The model’s architecture combines one-hour satellite mosaics with historical analysis in a single Functional Generative Network mesh transformer. It produces dense gridded fields, discrete cyclone tracks and native station-level predictions. The approach is notable because real-time observations are not merely used as a downstream correction; they are part of the model’s forecasting input.

Faster updates and smaller grid cells could be particularly useful when storms, fronts or precipitation systems develop quickly. More localized information may support agricultural scheduling and risk management, while wind, cloud and radiation estimates can help grid operators match renewable generation with demand. Google also positions the model as a way to expand access to high-resolution forecasting in parts of Latin America, Africa and the Asia-Pacific region, where regional numerical models can be expensive to run.

The available announcement does not provide a complete set of error metrics, precipitation scores or evaluation configurations. Its claims should therefore be read as Google’s description of the system, with independent testing still important for judging performance across regions and weather types. WeatherNext 3 is being integrated across Google Search, Gemini, Google Maps, Google Maps Platform and Google Cloud, extending AI weather forecasting from consumer queries into energy, agriculture and other operational settings.

Source: Google DeepMind