Google DeepMind's WeatherNext 3 delivers hourly satellite-fed forecasts at 5-km resolution

What Google announced

Google DeepMind and Google Research have unveiled WeatherNext 3, which the company calls its most advanced and accurate global AI weather model to date — a claim Google says is backed by independent live evaluations by Brightband. The model is being integrated across Search, Gemini, Maps, the Google Maps Platform and Google Cloud.

The headline change is what the model learns from. Instead of relying solely on output from traditional numerical weather prediction (NWP) systems, WeatherNext 3 ingests live observations, including a global mosaic of geostationary satellite imagery updated every hour.

Hourly forecasts at higher resolution

According to Google's announcement, most AI weather models — including the previous WeatherNext 2 — are trained on data from NWP models, physics-based supercomputer simulations that come with a six-hour data lag. That delay can introduce bias in fast-changing variables such as rain and surface temperature.

WeatherNext 3 combines live one-hour geostationary satellite mosaics with traditional historical analysis, allowing it to generate a new forecast every hour, each one grounded in the freshest observations available. Resolution is also up significantly: key surface variables such as temperature and moisture are produced on a 5-kilometre grid, other surface variables at 10 kilometres, and atmospheric variables like wind speed at 25 kilometres. Google puts the overall global picture at roughly five times sharper than WeatherNext 2, which produced forecasts on a 25-kilometre grid in six-hour increments.

The model also trains directly on sparse weather-station observation data rather than only on coarse atmospheric reanalyses, which Google says lets its global 5-kilometre forecasts account for regional topography such as coastlines, valleys and mountain ranges. The company frames this as particularly important for Latin America, Africa and Asia-Pacific, regions it says have historically lacked high-resolution forecasting because traditional regional models demand immense supercomputing resources.

Architecturally, the system feeds satellite mosaics and historical analysis into a single Functional Generative Network (FGN) mesh transformer that natively outputs dense gridded fields, discrete cyclone tracks and station-level predictions.

Better precipitation forecasting

Precipitation is a known weak spot for global models. Rain and snow systems are driven by fast-moving cloud processes on very small scales that physics-based simulations struggle to capture, and AI forecasts often produce blurry estimates or miss the boundaries of severe storms entirely.

Google says it addressed this by training WeatherNext 3 on two high-quality precipitation sources: NASA's satellite-based Integrated Multi-satellite Retrievals for GPM (IMERG) and Google's own global precipitation reanalysis built from satellite radar. Against medium-range baselines, the company reports Continuous Ranked Probability Score improvements of up to 60 percent versus IMERG, 30 percent versus MRMS, and 10 percent versus rain-gauge measurements at early lead times.

New clean-energy variables

WeatherNext 3 also introduces outputs aimed at renewable energy production. The model forecasts 100-metre wind speeds — roughly turbine height — for wind-energy output, alongside high-resolution cloud cover and solar radiation estimates to help solar farms gauge how much light will reach the ground. Google says this data lets grid operators and renewables developers predict generation from clean-energy assets and match it with consumer demand.

Why it matters

Weather drives decisions ranging from whether to carry an umbrella to how emergency services, farmers, air traffic controllers and supply chains operate. Two of WeatherNext 3's changes matter most in practice: an hourly update cycle anchored in live satellite data means fast-developing storms, fronts and precipitation systems can be caught earlier than a model on a six-hour lag allows, and native 5-kilometre global resolution extends high-fidelity forecasting to regions that supercomputer-based regional models have never economically served.

The rollout also shows how quickly AI forecasting has moved from research to default infrastructure: forecasts from WeatherNext 3 will surface in products used by billions of people, and the new energy variables tie weather modeling directly to grid planning and decarbonisation. One caveat is worth keeping in mind — the detailed accuracy figures come from Google's own evaluations against baselines, with the headline "most accurate" claim resting on Brightband's independent live scoring. How the model behaves against rivals such as ECMWF's AI systems in day-to-day use will be the real test.