WeatherNext 3: Google DeepMind moves weather AI off simulation-trained data

Google DeepMind and Google Research announced WeatherNext 3 on September 3, 2026 — the third generation of the weather-forecasting model family that started with GraphCast and has been running in production behind Google Search, Maps, and Pixel weather since WeatherNext 2. This isn't a research demo; it's a live upgrade to forecasts people already check every day, and it's worth understanding what actually changed under the hood.

What shipped

WeatherNext 3 generates a new global forecast every hour, at up to 5-kilometer resolution for surface variables like temperature and moisture, 10 kilometers for other surface fields, and 25 kilometers for atmospheric variables like wind speed. WeatherNext 2 forecast on a 25-kilometer grid, updated every 6 hours. DeepMind describes the jump as roughly five times sharper, and according to independent live evaluations by Brightband, it's currently the most accurate global weather model available.

The architecture is a single Functional Generative Network (FGN) mesh transformer that ingests live 1-hour geostationary satellite mosaics alongside historical analysis data, then outputs dense gridded fields, discrete cyclone tracks, and station-level point forecasts natively — three different output shapes from one model, rather than separate downstream systems bolted together.

The real change is the training data

The resolution numbers are the headline, but the more consequential decision is what the model learns from. Most AI weather models, WeatherNext 2 included, train on the output of numerical weather prediction (NWP) systems — physics simulations that are accurate but run on a roughly six-hour lag. That lag becomes a systematic bias for anything that moves fast: rain bands, temperature swings near coastlines, convective storms spinning up in an afternoon.

WeatherNext 3 instead trains directly on raw geostationary satellite mosaics and sparse weather-station observations. It's learning from the same class of data a forecaster would look at in real time, not from an intermediate simulation of it. That's a meaningfully different supervision signal, and it's the reason the model can refresh hourly instead of every six hours — it isn't waiting on a simulation pipeline to catch up.

Training on station observations directly also matters for geography that six-hour NWP simulations have historically underserved. High-resolution regional models are expensive to run, so large parts of Latin America, Africa, and Asia-Pacific have gotten coarser forecasts than North America or Europe. A globally consistent 5-kilometer model doesn't have that regional cost asymmetry baked in.

Precipitation and a new set of variables

Precipitation has always been the weak point for AI weather models — rain and snow are driven by cloud-scale processes that blur easily into diffuse, low-confidence probability fields. WeatherNext 3 trains on NASA's satellite-based IMERG precipitation product and DeepMind's own radar-based reanalysis, and the reported gains are large: up to 60% improvement in Continuous Ranked Probability Score against IMERG, 30% against MRMS radar, and 10% against rain-gauge measurements at early lead times.

The model also adds variables aimed squarely at grid operators rather than commuters: 100-meter wind speed at roughly turbine height, plus high-resolution cloud cover and surface solar radiation for solar-output estimation. Matching renewable generation to demand is a forecasting problem as much as an engineering one, and folding those variables into the same model that already produces the general forecast is a sensible consolidation rather than a separate product line.

Where it's already running

WeatherNext 3 is live now in Search, Gemini, Maps, Google Maps Platform, and Google Cloud. That distribution footprint is what makes this announcement different from a typical model paper: the accuracy gains land in products with billions of users on day one, not after a slow enterprise rollout.

My read: the resolution and speed numbers are the marketing headline, but the durable idea here is training weather models on observations instead of on another model's simulation output. That's the same pattern showing up across AI applications generally — closer supervision signal beats a cleaner but staler intermediate representation. Expect the next round of competition in this space to be less about mesh resolution and more about who has the best real-time observational data to train on.