Google DeepMind just dropped WeatherNext 3, and the numbers are genuinely startling. The model now forecasts down to 5km resolution — roughly three to five times finer than the 15-25km grids most AI weather models have been stuck on. Rain prediction improved 60% over WeatherNext 2, and it now produces hourly forecasts instead of the standard six-hour cadence. On Operational WeatherBench, the comparison tool built by startup Brightband, it outperformed models from Microsoft, Nvidia, ECMWF, and Google’s own previous work — plus traditional forecasts from the U.S. National Weather Service.
What’s different this time is the training strategy. The model has 2.4 times more parameters than its predecessor, but the real shift is what those parameters are pointed at. Researchers tuned decoder heads to target specific weather stations, not just the 3D grid averages most models output. Atmospheric scientist Daniel Rothenberg from Brightband put it well: predicting what Denver’s airport station will measure hourly “connects that forecasting task closer to the core.” DeepMind staff researcher Ferran Alet framed the philosophy bluntly — weather is chaotic, small differences perturb massively, and machine learning targets the actual problem: “approximate noisy physics from incomplete information and finite compute.”
There’s a catch, though. Google claims WeatherNext 3 is the “first” AI model to directly ingest raw satellite observations for high-resolution global forecasting. Startup WindBorne says its WeatherMesh 6 has been doing that with weather balloon data since late 2025. Google’s rebuttal is that its global resolution is higher, but the dispute highlights a broader truth: every one of these models still leans on national weather datasets at some stage. True end-to-end data assimilation remains unsolved.
The practical upshot arrives quickly — Google says WeatherNext 3 will feed forecasts into Search, Maps, and Gemini. That’s the part worth watching. The transformer revolution in meteorology has been as consequential as anything in LLMs, and unlike chatbot features, better weather prediction has measurable economic value. European agencies are already using AI models operationally. Bill Gates recently flagged AI weather forecasting as a tool for regions where supercomputers and sensor networks are unaffordable. If the 5km hourly resolution holds up outside research benchmarks, that promise gets a lot closer to real.
