Technology
Google releases WeatherNext 3 AI forecasting model with 5km resolution
The updated system uses real-time satellite data to provide hourly predictions and will integrate into Google Maps, Search, and Gemini.
The short version
- Google DeepMind and Google Research announced WeatherNext 3, an artificial intelligence model offering 5km-resolution hourly forecasts.
- The system is designed to improve precipitation tracking and will power weather features across Google Search, Google Maps, Gemini, and cloud platforms.
- Google claims the model leads benchmarks against traditional agency forecasts and competing deep-learning systems, though questions remain around priority claims for raw data assimilation.
Key facts
- Google DeepMind and Google Research developed WeatherNext 3, an AI weather model featuring 2.4 times more parameters than its predecessor.[TechCrunch]
- The model provides hourly forecasts at a 5-kilometer resolution and achieves a 60% improvement in rain evaluations compared to WeatherNext 2.[TechCrunch]
- Google plans to integrate WeatherNext 3 outputs into Google Search, Google Maps, Gemini, and its enterprise cloud tools.[TechCrunch]
- Testing on Brightband's Operational WeatherBench indicated WeatherNext 3 outperformed competing models from Microsoft, Nvidia, and the ECMWF, as well as conventional supercomputer-driven forecasts.[TechCrunch]
- The model trains directly on real-time hourly satellite observations and ground weather stations in addition to government meteorological datasets.[TechCrunch · The Verge]
What remains uncertain
- Startup WindBorne disputed Google's claim of being the first AI model to ingest raw observations directly, stating its WeatherMesh 6 model has done so since late 2025.[TechCrunch]
- Both Google and competing AI weather models still depend in part on pre-processed national datasets, leaving complete direct data assimilation an ongoing development goal.[TechCrunch]