Introducing WeatherNext 3, our most advanced and accurate global weather AI model
2026-09-04 · Google DeepMind
WeatherNext 3: Google's Next Leap in AI Weather Forecasting
On September 3, 2026, Google DeepMind and Google Research introduced WeatherNext 3, the latest generation of Google's flagship AI weather forecasting model. According to independent live evaluations by Brightband, it ranks as the most advanced and accurate global weather model to date.
Why This Matters
Weather shapes billions of decisions every day — from whether to carry an umbrella to how heatwaves and droughts ripple through agriculture, energy grids, supply chains, and national economies. AI has already transformed forecasting by enabling faster and more accurate predictions than traditional physics-based methods. Even so, capturing highly local and rapidly changing conditions has remained a hard problem. Previous models lacked sufficient spatial resolution and struggled to incorporate real-time satellite observations.
The Core Advance: Learning From Live Satellite Data
The most significant change in WeatherNext 3 is what it learns from.
Most previous AI weather models, including WeatherNext 2, trained on output from numerical weather prediction (NWP) systems — supercomputer-driven physics simulations that carry a six-hour data lag. That lag introduces bias for fast-changing variables like precipitation and surface temperature.
WeatherNext 3 instead ingests a continuously updated mosaic of live global geostationary satellite imagery, giving it a real-time view of the atmosphere. This allows the model to:
- Generate a fresh forecast every hour, each grounded in the latest satellite observations
- Detect rapidly developing storms, fronts, and precipitation systems at an earlier stage
- Reduce bias that stems from stale NWP-derived training data
Five Times Sharper Resolution
WeatherNext 3 produces hourly forecasts at multiple spatial resolutions while maintaining physical consistency from planetary wind patterns down to local topography:
| Variable type | Resolution |
|---|---|
| Key surface variables (temperature, moisture) | 5 km |
| Other surface variables | 10 km |
| Atmospheric variables (wind speed) | 25 km |
Overall, this is roughly five times sharper than WeatherNext 2, which operated on a 25-kilometer grid at six-hour intervals. A side-by-side comparison over the UK shows WeatherNext 3 resolving intricate local terrain clearly, eliminating the pixelated, over-smoothed thermal patterns visible in older model output.
Architecture: A Single FGN Mesh Transformer
WeatherNext 3 runs on a single, flexible Functional Generative Network (FGN) mesh transformer. It ingests one-hour live geostationary satellite mosaics alongside traditional historical analysis data and produces three output types:
1. Dense gridded fields — continuous spatial variables such as temperature, humidity, and wind
2. Discrete cyclone tracks — dedicated tropical cyclone path prediction
3. Station-level sparse coordinate predictions — precise forecasts tied to specific observation points
Applications and Integration
The improvements carry the most weight in these areas:
- Agriculture: Better precipitation and temperature forecasts support irrigation scheduling and crop management decisions
- Clean energy: New variables for wind and solar output improve grid dispatch planning
- Extreme weather response: Earlier, higher-resolution storm warnings give emergency services more lead time
- Daily planning: Hourly localized forecasts give individuals more reliable guidance for outdoor activities and travel
WeatherNext 3 is now integrated across:
- Google Search and Google Maps — enhanced weather information for everyday users
- Gemini — natural-language weather queries powered by the new model
- Google Maps Platform — high-resolution weather data available to developers
- Google Cloud — enterprise integration for custom applications and workflows
Takeaway
WeatherNext 3 represents a genuine architectural shift: by learning directly from real-world satellite observations rather than processed NWP output, it closes the data latency gap and delivers more timely, localized, and accurate forecasts. With the rollout spanning Google's global product ecosystem, the improvements reach billions of users directly.