Google DeepMind releases WeatherNext 3, its most accurate AI weather model yet
The AMW Read
Incremental model-line upgrade from an established foundation-model lab that meaningfully expands distribution into core Google consumer products, but does not meet any cross-substrate capital, compute, or talent threshold.
Google DeepMind releases WeatherNext 3, its most accurate AI weather model yet
Google DeepMind and Google Research released WeatherNext 3, a new AI weather forecasting model that Google says will begin feeding core weather data into Search, Google Maps, and Gemini, in addition to being available to researchers and developers on Google Cloud. The model topped Operational WeatherBench, an independent benchmark built by startup Brightband, beating deep-learning rivals from Microsoft, Nvidia, and the European Centre for Medium-Range Weather Forecasting, as well as traditional physics-based forecasts from the U.S. National Weather Service and ECMWF. WeatherNext 3 has 2.4 times more parameters than its predecessor, resolves key variables down to 5 km versus the 15-25 km typical of earlier AI models, improves rainfall accuracy 60% over WeatherNext 2, and produces hourly forecasts instead of six-hourly ones by ingesting real-time satellite data rather than relying solely on government-processed datasets.
The upgrade matters because it pushes AI weather models past matching legacy supercomputer forecasts and toward directly powering consumer products at Google's scale — the company says this is the first time these variables will feed its products directly. It also sharpens a competitive field that includes WindBorne's WeatherMesh 6, which the AI weather startup says has ingested raw observations from its own weather-balloon fleet since late 2025; Google's counter is higher global resolution, not primacy on raw-data ingestion. The broader signal is that transformer-based forecasting is now displacing government supercomputers on accuracy, not just speed, in a domain where compute cost has historically gated access to good forecasts.
For builders, WeatherNext 3's cloud availability lowers the barrier for downstream products in logistics, agriculture, insurance, and disaster response to build on frontier-grade forecasts without owning forecasting infrastructure. For investors tracking AI weather startups like Brightband and WindBorne, Google folding a state-of-the-art model directly into Search, Maps, and Gemini raises the distribution bar those startups must clear to stay differentiated.
