WeatherNext 2: Our most advanced weather forecasting model

Source: DeepMind Date: 2025-11-17 URL: https://deepmind.google/blog/weathernext-2-our-most-advanced-weather-forecasting-model/

Summary

Google DeepMind released WeatherNext 2, a Functional Generative Network architecture that surpasses its predecessor on 99.9% of evaluated atmospheric variables while running 8x faster on a single TPU, with forecast resolution down to hourly intervals at 15-day range. The key innovation is learning from individual weather element distributions (marginals) to accurately predict complex system interdependencies (joints). Available via Earth Engine, BigQuery, Vertex AI early access, and integrated into Google Search, Gemini, Pixel Weather, and Maps.

Implications

99.9% improvement across variables with 8x speedup is a compound win. Quality improvements alone are expected in new model versions; quality + dramatic compute efficiency together is the combination that matters for production deployment. A single TPU generating hundreds of scenarios in under a minute is deployable at the scale of a consumer product, not just a research cluster.

Hourly resolution at 15-day range changes the forecast granularity game. Traditional numerical weather prediction at that range is inherently uncertain; AI models generating hourly-granularity scenarios up to 15 days out is a fundamentally different product for energy trading, aviation, agriculture, and logistics.

Google’s weather distribution stack is now complete. Search, Gemini, Pixel, Maps, Earth Engine, BigQuery, Vertex AI — WeatherNext 2 data is feeding every Google consumer and enterprise surface. That’s a vertical integration of weather intelligence that no competitor can match (Meteomatics, ECMWF, and The Weather Channel are all specialized, not horizontally integrated).

Watch:

• CRPS performance against ECMWF ENS and the new DeepMind cyclone model — is WeatherNext 2 the better general model or the cyclone model fills a specialized niche?
• Vertex AI early access conversion to GA timeline
• Energy and commodities trading adoption — hourly 15-day forecasts at this speed are exactly what systematic trading strategies need