Five Kilometers, Every Hour: How Google DeepMind's WeatherNext 3 Rewrites Global Forecasting

On September 3, 2026, Google DeepMind and Google Research introduced WeatherNext 3, which the company describes as its most advanced and accurate global weather model to date — a claim it backs not with internal benchmarks but with independent live evaluations by Brightband, the third-party scoring service that runs continuous head-to-head comparisons of operational forecasting systems. The launch is less a single model refresh than a top-to-bottom redesign of how an AI forecaster sees the atmosphere, and it is already powering the weather surfaces billions of people touch every day, including Google Search, the Gemini app, Google Maps, the Maps Platform Weather API, and Earth Engine.

The resolution problem, solved from two directions

A forecast is only as useful as its detail. WeatherNext 2, the previous generation, produced global forecasts on a 25-kilometer grid in 6-hour increments — adequate for synoptic trends, but blind to the weather that actually decides whether you get wet. WeatherNext 3 attacks that limitation on both the time and space axes simultaneously: it generates a new forecast every hour, and it renders key surface variables like temperature and moisture at a native 5-kilometer (0.05°) resolution, with other surface variables at 10 kilometers and atmospheric variables such as wind speed at 25 kilometers. The result is a global weather picture roughly five times sharper than its predecessor.

The difference is visible in the company’s own comparison imagery: 2-meter temperature forecasts over the United Kingdom that once looked like pixelated, over-smoothed color blocks now resolve the intricate thermal texture created by coastlines, valleys, and hill country. For communities near complex topography — where temperature and humidity can swing dramatically across just a few kilometers — that detail is the difference between a forecast that describes a region and one that describes your neighborhood.

Learning from satellites and stations, not simulations

The deeper architectural change is what the model learns from. Most AI weather models, including WeatherNext 2, are trained on outputs from numerical weather prediction (NWP) systems — supercomputer-driven physics simulations that are powerful but carry a six-hour data lag. That lag systematically biases fast-changing variables like rain and surface temperature, because the model’s picture of “now” is always half a day old.

WeatherNext 3 breaks that dependency. It ingests a live, global mosaic of 1-hour geostationary satellite observations as a direct model input, alongside traditional historical analysis, feeding a single flexible Functional Generative Network (FGN) mesh transformer. The architecture outputs dense gridded forecast fields, discrete cyclone tracks, and — natively — predictions at sparse, station-level coordinates. According to the accompanying paper, the model has roughly 2.4 times the parameters of WeatherNext 2.

Just as significantly, the model trains directly on sparse weather station observation data rather than only on gridded reanalysis. Because weather stations sit where people actually live — in valleys, on coastlines, on mountain slopes — the 5-kilometer global grid inherits real regional detail instead of interpolating it. Google frames this as an equity story as much as a technical one: regions across Latin America, Africa, and Asia-Pacific have historically been underserved by high-resolution forecasting because regional NWP modeling at that fidelity requires supercomputing budgets few national meteorological services can afford. A single global model that resolves local topography by default brings high-fidelity forecasting to billions of people who never had it.

Fixing the hardest problem: precipitation

Precipitation has always been the graveyard of weather models, AI and physics-based alike. Rain and snow are driven by fast-moving convective processes on tiny spatial scales, which is why AI precipitation forecasts historically come out blurry or miss storm boundaries entirely — a failure mode that matters when the boundary is a severe thunderstorm.

WeatherNext 3 takes a data-first approach: it trains on two exceptionally 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. The payoff, measured by Continuous Ranked Probability Score (CRPS) in medium-range global forecasts, is an improvement of up to 60% against IMERG baselines, 30% against MRMS, and 10% against rain gauge measurements at early lead times. In Google’s consumer products, users planning a day or more ahead will see precipitation forecasts up to 50% more accurate, with the largest gains concentrated in regions where forecasts have historically been least reliable.

Built for the grid, not just the umbrella

The most commercially significant novelty may be the clean-energy variable set. WeatherNext 3 forecasts 100-meter wind speeds — approximately turbine height — for wind-farm output estimation, alongside high-resolution cloud cover and surface solar radiation for solar producers. Grid operators and renewables developers can now plan how much power their assets will generate and match it against demand using the same hourly-cadence model that powers a Maps widget. As renewable penetration rises, forecast error in wind and solar output translates directly into balancing costs and fossil standby generation; a sharper, faster-updating global model is infrastructure in the fullest sense.

Access extends beyond Google’s own apps. Hourly-updated global forecast data is queryable in BigQuery and Earth Engine and bulk-downloadable from Google Cloud Storage, with no model setup required — positioning WeatherNext 3 as a data substrate for researchers, insurers, logistics firms, and agricultural platforms.

The quiet stakes

It is tempting to file WeatherNext 3 under incremental product polish. The specifics argue otherwise. An hourly update cycle fed by live satellite data compresses the warning time available for rapidly materializing storms, fronts, and precipitation systems. Station-native training collapses a decades-long gap in forecast quality between the wealthy mid-latitudes and the Global South. And the renewable energy variables arrive exactly as grids worldwide wrestle with the variability of wind and solar at unprecedented scale.

The atmosphere will always retain a degree of unpredictability, as Google’s own team is careful to note — and the company directs users to national meteorological agencies for official severe-weather warnings. But by training on real-world observations rather than simulations of them, WeatherNext 3 moves global forecasting measurably closer to matching what is actually happening on the ground, one square five-kilometer cell at a time.