Our WeatherNext 2 AI model demonstrated a massive leap forward in predicting cyclones.

Source: Google Date: 2026-08-06 URL: https://blog.google/innovation-and-ai/models-and-research/google-deepmind/weathernext-2-cyclones/

Summary

Google DeepMind’s WeatherNext 2 model achieved state-of-the-art accuracy predicting a cyclone’s track, intensity, and wind structure — described as roughly a decade of meteorological progress in one model, published in Nature, and open-sourced for the research community. The gain translates to about an extra day of reliable warning time ahead of a storm.

Implications

A model-landscape signal outside the LLM lane: it’s a reminder that the “frontier model” race extends well beyond chat/agent models into domain-specific scientific systems, and that Google is treating open-sourcing as a credibility move here (as with prior DeepMind science releases) even while keeping its flagship Gemini weights closed. Worth a line in any model-landscape roundup that tracks compute allocation across domains, not just LLM parameter counts.