Google DeepMind WeatherNext 3 Officially Released: A New Milestone in AI Weather Forecasting

Introduction: Redefining Forecast Precision and Timeliness

On September 3, 2026, Google DeepMind and Google Research jointly released WeatherNext 3, the latest version of Google's flagship operational AI weather model. WeatherNext 3 achieves major breakthroughs in technical architecture, forecast accuracy, and data integration, marking a new development phase for AI-driven weather forecasting.

The core innovation of WeatherNext 3 is its hourly update capability. Traditional numerical weather prediction models typically carry a six-hour data assimilation lag, but WeatherNext 3 eliminates this bottleneck by directly ingesting live geostationary satellite imagery, dramatically improving forecast timeliness.

Technical Architecture: Functional Generative Network and Mesh Transformer

WeatherNext 3 employs a new Functional Generative Network (FGN) with a mesh transformer architecture, representing the core technical differentiation from previous models.

Direct Satellite Data Ingestion

Unlike traditional methods relying on reanalysis data — which inherently carries a six-hour lag from physical models — WeatherNext 3 directly ingests live geostationary satellite mosaics. This allows forecasts to update every hour, significantly improving responsiveness to rapidly developing weather systems.

Multi-Resolution Output

WeatherNext 3 operates at variable resolutions depending on output type:

• 5 km: Surface temperature and dew point (station-trained)
• 10 km: General gridded surface fields including precipitation, wind, clouds, and solar radiation
• 25 km: Atmospheric variables across 13 pressure levels

64-Member Ensemble Forecasting

WeatherNext 3 provides 64-member probabilistic ensemble forecasts covering 15 days globally, offering uncertainty quantification rather than single deterministic predictions.

Forecast Accuracy: Major Improvements Across Key Metrics

Google reports significant improvements across multiple key metrics:

Precipitation Forecasting

• Up to 50% more accurate for forecasts a day or more ahead
• Up to 60% improvement in CRPS against NASA IMERG satellite data
• 30% improvement against MRMS radar
• 10% improvement against rain gauges

Temperature Forecasting

• Up to 30% improvement in station-level temperature vs. WeatherNext 2
• Up to 40% better than ECMWF ensemble at short lead times

Tropical Cyclone Tracking

• Approximately 10% average advantage over ECMWF AIFS ENS v2 in the first forecast week

These figures indicate that WeatherNext 3 matches or exceeds traditional top-tier numerical weather prediction systems across multiple key scenarios.

Integration and Availability

WeatherNext 3 is integrated into multiple Google consumer and developer products:

Consumer Products

• Google Search: More accurate forecasts when users search for weather
• Google Maps: Updated weather information in route planning and navigation
• Gemini App: Detailed weather analysis through conversational interface

Developer and Enterprise

• Google Maps Platform Weather API: High-accuracy weather data for third-party applications
• Google Cloud (Zarr format): Large-scale data analysis for researchers and enterprises
• Earth Engine: Earth science research and environmental monitoring
• BigQuery: Integrated analysis of weather and business data

Notably, WeatherNext 3 is not open source. Researchers and businesses must apply for dataset access through Google Cloud, with an approval window of five to seven business days.

Special Significance for Asia-Pacific

WeatherNext 3 holds particular importance for the Asia-Pacific region:

Typhoon Forecasting: The Asia-Pacific is one of the world's most typhoon-active regions. The 10% accuracy improvement in tropical cyclone tracking has direct implications for disaster prevention in Hong Kong, Taiwan, the Philippines, Japan, and other typhoon-prone areas.

Agricultural Meteorology: Major agricultural nations including Southeast Asia, China, and India have urgent needs for accurate precipitation forecasts. The 50% precipitation accuracy improvement supports agricultural production decisions, irrigation management, and food security.

Urban Heat Islands: The 5km resolution temperature forecasts provide valuable reference for urban planning and energy management in high-density cities like Hong Kong, Singapore, and Shanghai.

Shipping and Logistics: The Asia-Pacific is the world's busiest shipping region, where accurate marine weather forecasts are critical for route planning and port operations.

Usage Limitations and Considerations

Google emphasizes that WeatherNext 3 is intended for planning and guidance, not as a replacement for official safety services:

• Not a Warning System: Official severe weather warnings remain the responsibility of national meteorological agencies
• Live Radar Superiority: For immediate 30-90 minute precipitation tracking, live radar remains more reliable than AI model predictions
• Training Data Limitations: As an AI model trained on historical data, WeatherNext 3 may face challenges predicting unprecedented extreme weather events outside its training distribution

Conclusion: The Future of AI Weather Forecasting

WeatherNext 3's release represents an important milestone in AI weather forecasting technology. Through hourly updates, multi-resolution output, and direct satellite data ingestion, WeatherNext 3 surpasses traditional numerical weather prediction models across multiple timeliness and accuracy metrics.

As WeatherNext 3 integrates into Google's consumer and enterprise product ecosystem, AI-driven weather forecasting will reach broader user groups, providing more accurate decision support for agriculture, shipping, energy, insurance, and urban management. For the Asia-Pacific region, this technological breakthrough holds particularly important application potential in typhoon forecasting, agricultural meteorology, and urban planning.