TL;DR

Google DeepMind has published WeatherNext results in Nature and released the model weights and code openly. A three-day WeatherNext forecast is as accurate as what earlier systems managed at two days — roughly an extra day of warning on where a cyclone goes, how strong it gets and how its winds are structured. The UK Met Office is named among the collaborating forecasting agencies, alongside the US National Hurricane Center and CIRA.

Solving both halves of the problem at once

Cyclone forecasting has historically forced a choice between two modelling approaches. Track is governed by large-scale atmospheric currents, best captured by coarse global models. Intensity comes from fine-grained thermodynamics near the storm’s core, which has required specialised high-resolution local models. WeatherNext is a single system predicting both.

Training drew on just under 20 terabytes of worldwide atmospheric readings plus the IBTrACS historical database, which covers close to 5,000 storms. Functional Generative Networks produce ensembles cheaply enough that a 15-day forecast takes under a minute on a single TPU. Last year the system generated 50 predictions at a time, matching physics-based global models; this year it runs 1,000, which is what allows rare tail-risk scenarios such as rapid intensification to surface at all.

The result the team flags as genuinely unexplained is resolution. WeatherNext Cyclones operates on 28km data — a hundred times coarser than conventional models — and a smaller variant at 111km still performs well. Why accuracy survives that reduction remains, in their words, an open research question.

The open release is the notable part

DeepMind is publishing weights and code for WeatherNext 2, WeatherNext Cyclones and a compact WeatherNext 2-mini that runs in a free Colab notebook. That is a meaningful contrast with the direction of travel elsewhere this week, as Alibaba and Moonshot move to extract revenue shares from heavy commercial users of their open-weight releases. Public-interest forecasting has different economics from a commercial model licence, but the choice is still a choice.

The operational record is not hypothetical. During the 2025 hurricane season the model helped the National Hurricane Center anticipate Hurricane Melissa’s rapid intensification and Jamaican landfall, buying preparation time on the ground.

Looking forward

For the UK, the Met Office’s involvement gives national forecasting direct access to the method rather than the output, and an open release means smaller meteorological agencies and academic groups can build on it. DeepMind’s own caveat stands: official warnings still come from national weather services, not from a model on a laptop.