startupfortune.com
Google DeepMind Open-Sources WeatherNext 2 for Faster Hurricane Forecasts
Google DeepMind has open-sourced WeatherNext 2, an AI model that forecasts hurricanes far faster than physics-based systems and beats its own predecessor on nearly every variable it tracks. Google DeepMind this week released the code and weights behind WeatherNext 2 and a companion cyclone model, WeatherNext Cyclones, on GitHub, as Engadget reported. Anyone, a university lab, an airline, an insurer, a government forecasting office, can now download and run the same system Google uses to power weather results in Search, Gemini, Pixel Weather and its Maps Platform Weather API. That's not a paper. That's the model itself. The headline number is speed. WeatherNext 2 completes a full ensemble forecast, hundreds of possible weather scenarios generated from noise injected into a single input, in under a minute, running on just one TPU. Google says that's roughly eight times faster than the physics-based models national weather agencies have relied on for decades, and it comes at up to one-hour resolution, sharper than most operational systems produce. Accuracy improved too. According to DeepMind, WeatherNext 2 beats its prior generation, built on the GenCast architecture, on the large majority of the variables it tracks, temperature, wind, humidity, across every lead time from zero to...