Google DeepMind and Google Research have released a new global weather forecasting model, called WeatherNext 3, that trains directly on live satellite observations instead of relying only on the output of conventional physics-based simulations. The shift lets the system refresh its predictions every hour and resolve conditions as fine as 5 kilometers, a jump the companies say meaningfully sharpens forecasts for rain, wind and other fast-changing weather.
The model went live on September 3, 2026, and is already built into Google Search, the Gemini app, Google Maps and the Google Maps Platform Weather API, with raw forecast data available to developers through Google Earth Engine, BigQuery and Cloud Storage. According to independent live evaluations run by the forecasting benchmark group Brightband, WeatherNext 3 currently ranks as the most accurate global weather model being tracked.
Weather influences an enormous range of daily decisions, from whether to carry an umbrella to how utilities balance electricity demand during a heat wave, and Google frames the update as useful to a wide range of specialists: an emergency responder tracking a sudden wind shift, an air traffic controller adjusting flight paths around a storm cell, or a farmer deciding when to plant or harvest. The company also runs a public visualization tool called Weather Lab, where the new model’s output can be viewed in near real time alongside comparisons to earlier forecasting systems.
Training on live satellite mosaics instead of six-hour-old model output
Most AI weather systems, including the earlier WeatherNext 2, learn from numerical weather prediction (NWP) data: supercomputer-driven physics simulations that carry roughly a six-hour lag before their output becomes available. That delay can distort fast-moving variables such as surface temperature or rainfall bands. WeatherNext 3 instead ingests a continuously updating mosaic of live geostationary satellite imagery, producing a new global forecast every hour, each one grounded in the most recent observations rather than a stale simulation run.
The system also trains directly on sparse weather station observations rather than smoothed atmospheric representations, which helps it account for local terrain effects near coastlines, valleys and mountain ranges that older models tend to blur together.
Five-kilometer resolution replaces a 25-kilometer grid
WeatherNext 2 produced forecasts on a 25-kilometer grid updated in 6-hour increments. WeatherNext 3 resolves surface variables like temperature and moisture at 5 kilometers, other surface variables at 10 kilometers, and atmospheric variables such as wind speed at 25 kilometers, while keeping the physics consistent from broad global circulation patterns down to local topography. Google describes the overall picture as roughly five times sharper than its predecessor. The architecture behind it, a Functional Generative Network mesh transformer described in a paper posted to arXiv, outputs dense gridded fields, discrete cyclone tracks and station-level forecasts from a single model rather than separate specialized systems.
Closing the precipitation gap with satellite radar data
Rain and snow have historically been the hardest quantities for any weather model to pin down, since the cloud processes that drive them operate on scales too small for physics simulations to capture cleanly, often producing blurred estimates that miss a storm’s real boundary. To address that, WeatherNext 3 trains on two satellite-derived precipitation datasets: NASA’s Integrated Multi-satellite Retrievals for GPM (IMERG) and Google’s own global precipitation reanalysis built from satellite radar. In medium-range evaluations, Google reports Continuous Ranked Probability Score improvements of up to 60% against IMERG, 30% against the MRMS radar network, and 10% against direct rain gauge readings at early lead times. For forecasts a day or more out, the company says precipitation accuracy improves by as much as 50%, with the largest gains in regions that have historically had the least reliable coverage. Rain and snow have long been the weak point of global weather modeling because the cloud processes that generate them operate on scales far smaller than the grid a typical physics simulation can resolve, which tends to produce blurred rainfall estimates that miss the sharp edges of a real storm front. Training on satellite radar and ground-truth precipitation data, rather than a coarser simulated approximation of rainfall, is what lets WeatherNext 3 recreate those sharper boundaries.
New forecasts built for wind and solar power planning
Beyond day-to-day forecasts, WeatherNext 3 adds outputs aimed squarely at clean energy operators. It predicts wind speed at 100 meters, roughly turbine height, to help estimate wind-energy output, along with high-resolution cloud cover and surface solar radiation so solar farms can project how much sunlight they will receive. Grid operators and renewable developers can use those figures to match expected generation against consumer demand more precisely than was previously possible with coarser regional models.
Reaching regions that lacked supercomputer-scale forecasting
Traditional high-resolution regional weather models require enormous supercomputing budgets, which has left large parts of Latin America, Africa and the Asia-Pacific region without the kind of localized forecasting available in wealthier countries. Because WeatherNext 3 generates its 5-kilometer global forecasts from a single model rather than a patchwork of regional systems, Google says the same fidelity now extends to those underserved areas without requiring local supercomputing infrastructure.
Availability across Search, Maps and Google Cloud
Google is rolling the model out simultaneously across consumer and developer products: forecasts in Search, Gemini and Maps for everyday planning, and raw hourly data in BigQuery, Earth Engine and Cloud Storage for researchers and businesses that want to build their own applications on top of it. The company frames the update as a step toward forecasts that track what is actually happening in the atmosphere in near real time, rather than reconstructing it from a simulation that is already hours old, though it continues to point users to national meteorological agencies for official severe-weather warnings.
This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.
More from Morning Overview
