WeatherNext 3 Delivers 5x Sharper Forecasts via Live Satellite Data

Google DeepMind and Google Research shipped WeatherNext 3 on September 3, 2026, claiming the title of most advanced and accurate global weather model based on independent live evaluations by Brightband. The release is technically significant not because it adds parameters to an existing pipeline, but because it replaces the training data substrate entirely — swapping physics-simulation output for raw, real-time geostationary satellite observations — and adopts an architecture purpose-built to consume that stream.

For engineers tracking where domain-specific AI is heading, this is a concrete demonstration that architectural specificity outperforms GPU scaling: the accuracy gains come from a fundamentally different data ingestion path and mesh-transformer design, not from scaling a conventional numerical weather prediction (NWP) surrogate.

Architecture: FGN Mesh Transformer on Live Satellite Data

WeatherNext 3 is built around a single Functional Generative Network (FGN) mesh transformer. The system ingests live 1-hour geostationary satellite mosaics alongside traditional historical analysis data, then outputs three forecast representations simultaneously: dense gridded fields, discrete cyclone tracks, and station-level sparse coordinate predictions.

The sparse-coordinate output is particularly notable. Rather than mapping atmospheric state onto a fixed analysis grid, the model trains directly on sparse weather station observation data, resolving sub-grid phenomena — coastal thermal gradients, valley inversions, orographic precipitation — that conventional analysis grids erase.

Previous AI weather models, including WeatherNext 2, were trained on NWP model output. NWP pipelines carry an inherent six-hour data lag. By training on continuously updating satellite mosaics, WeatherNext 3 issues a new forecast every hour, each conditioned on the most recent satellite state.

Resolution and Variable Coverage

WeatherNext 3 operates at three spatial resolutions. Key surface variables — temperature and moisture — resolve at 5 kilometers (0.05°). Other surface variables run at 10 kilometers. Atmospheric variables such as wind speed output at 25 kilometers. WeatherNext 2 produced all forecasts on a uniform 25-kilometer (0.25°) grid in 6-hour increments. The net effect is a global weather picture roughly five times sharper than its predecessor, with forecast cadence improving from every 6 hours to hourly.

The model also introduces two variable families absent from WeatherNext 2: 100-meter wind speed forecasts (roughly turbine-hub height) and high-resolution cloud cover with surface solar irradiance — both engineered for renewable energy dispatch planning.

Precipitation Benchmarks

Precipitation has historically been the hardest variable for global AI models because convective processes operate below typical model grids and cloud microphysics evolve faster than assimilation cycles can track. WeatherNext 3 trains on two high-quality precipitation datasets: NASA's Integrated Multi-satellite Retrievals for GPM (IMERG) and a proprietary global precipitation reanalysis derived from satellite radar.

In medium-range global evaluations, the model achieves a Continuous Ranked Probability Score (CRPS) improvement of up to 60% relative to IMERG, 30% relative to MRMS, and 10% against rain gauge measurements at early lead times. Precipitation output is produced at 11-kilometer resolution — compared to 25 kilometers in WeatherNext 2 — which closes the gap between the model's probability-of-precipitation footprint and satellite ground truth on convective band structure.

For end users accessing forecasts through Google Search, Gemini, or Google Maps, the practical result is up to 50% more accurate precipitation forecasts for day-and-beyond planning horizons, with the largest gains in Latin America, Africa, and Asia-Pacific, where supercomputing costs have historically prevented high-resolution regional NWP coverage.

Availability and Data Access

WeatherNext 3 began rolling out on September 3, 2026, across Google Search, the Gemini app, Google Maps, Google Maps Platform's Weather API, and Google Earth Engine. For programmatic access, forecast data is queryable in BigQuery and Earth Engine, with bulk download available from Google Cloud Storage — no model deployment required on the consumer side.

WeatherNext 3 demonstrates that committing fully to a domain-native data source — raw satellite radiance rather than NWP-processed analysis fields — and designing an architecture around that source's temporal and spatial properties produces accuracy gains that grid refinement on legacy pipelines cannot replicate. As renewable energy grids demand tighter generation forecasts and climate volatility raises the operational cost of forecast errors, the appetite for this class of specialized physical-system AI will grow.