Tropical cyclones have killed more than 700,000 people and caused $1.4 trillion in economic losses over the past 50 years. Every extra hour of warning matters. Now Google DeepMind says its WeatherNext AI model gives forecasters a full additional day of predictive accuracy on cyclone tracks, intensity, and wind structure — an improvement the team equates to roughly a decade of traditional meteorological progress. And they’ve open-sourced the whole thing.
What WeatherNext Actually Does
The model, described in a paper published in Nature, predicts a cyclone’s track, intensity, and wind structure up to 15 days in advance. The key claim: WeatherNext’s three-day forecasts are as accurate as what previous models could deliver for only the next two days. That’s 24 hours of additional lead time — enough to move people, stage supplies, and issue warnings that arrive before the storm, not during it.
WeatherNext bridges a gap that has split cyclone forecasting for years. A cyclone’s track is driven by massive global atmospheric currents, best modelled by coarse global models. Its intensity is driven by fine-scale thermodynamic processes around the eye wall, best modelled by high-resolution local models. WeatherNext handles both in a single AI model, trained end-to-end on nearly 20 terabytes of global atmospheric data and the IBTrACS database of nearly 5,000 historical storms.
What’s surprising is the resolution. WeatherNext Cyclones operates at 28×28 km — 100 times coarser than traditional intensity models. A compact version, WeatherNext 2-mini, runs at 111×111 km and still performs well. DeepMind says it remains an “open research question” how the model achieves such accuracy at this resolution.
Real-World Impact Already
This isn’t theoretical. During the 2025 hurricane season, WeatherNext helped the National Hurricane Center predict Hurricane Melissa’s rapid intensification and landfall in Jamaica — a historic forecast that gave teams on the ground critical preparation time. This year, the system generates 1,000 possible scenarios for each cyclone, up from 50 last year, capturing rare but devastating events like rapid intensification.
A single 15-day forecast runs in under a minute on a TPU. That speed matters: forecasters can rapidly evaluate probability distributions of worst-case scenarios instead of waiting hours for a single physics-based simulation.
The model was co-developed with the National Hurricane Center, the Cooperative Institute for Research in the Atmosphere (CIRA), and the UK Met Office. It’s part of Google Earth AI, and forecasts are visualised on the Weather Lab platform.
Open Source — With a Compact Version Anyone Can Run
DeepMind has open-sourced the code and model weights for both WeatherNext 2 and WeatherNext Cyclones. A compact version, WeatherNext 2-mini, runs on a single TPU in a free public Colab notebook.
This is notable because it’s not just a research paper with a GitHub link that goes nowhere. The model weights, the inference code, and a working notebook are all available. Meteorological agencies, researchers, and nonprofits can build on it without a Google partnership.
The open-source release puts DeepMind’s approach in contrast to the broader trend of frontier labs keeping their most capable models behind API walls. As we’ve covered with China’s open-weights strategy and Anthropic’s open-weights position, the open-versus-closed debate is usually about language models. WeatherNext shows the calculus can be different for scientific AI — where the upside of broad adoption directly saves lives.
What This Means for the Pacific
For New Zealand and the broader Pacific, cyclone forecasting is not abstract. The South Pacific cyclone season runs November to April. Fiji, Vanuatu, Tonga, and the Cook Islands sit directly in cyclone paths. NIWA’s seasonal climate outlooks and MetService’s warnings depend on the same global modelling infrastructure that WeatherNext now outperforms.
The open-source release means Pacific meteorological services could, in principle, run WeatherNext locally or via cloud TPUs without licensing fees. The practical barrier is compute access — a single TPU is not cheap — but the Colab notebook demo suggests the mini model can run on accessible infrastructure. For countries that have historically relied on delayed forecasts from Northern Hemisphere agencies, direct access to a state-of-the-art cyclone model is a meaningful shift.
The tension between AI’s energy footprint and its climate applications is worth noting here. WeatherNext is an example of AI doing something about climate rather than just consuming energy to talk. The compute cost of running cyclone forecasts is trivial compared to the economic and human cost of an unwarned landfall.
