DeepMind opens WeatherNext, buys forecasters a day on cyclones

Google DeepMind put the code and weights for its cyclone forecasting stack on GitHub this week, and the headline claim is a familiar-sounding but genuinely useful one. According to DeepMind's blog post, the model gives forecasters an extra day of predictive accuracy on average, with three-day forecasts as good as what prior models could manage at two days. The company frames that gain as roughly a decade's worth of meteorological progress compressed into one release.

Three variants are going out: WeatherNext Cyclones, WeatherNext 2, and WeatherNext 2-mini, the last of which operates at a coarser 111x111 km resolution and is small enough to run on a single TPU via a free Colab notebook. The team says a single 15-day forecast runs in under a minute on a TPU, and that ensembles have been scaled from 50 members last year to 1,000 this year, which matters because the tail scenarios (rapid intensification, unexpected landfalls) are precisely what forecasters need to plan for. The code and weights are on GitHub.

DeepMind points to the 2025 hurricane season for a real-world case, saying the model helped the National Hurricane Center make a historic forecast for Hurricane Melissa by predicting the storm's rapid intensification and landfall in Jamaica. The stakes framing in the post is worth quoting on: tropical cyclones are responsible for more than 700,000 deaths and $1.4 trillion in economic losses globally over the past 50 years, so even a modest extension of lead time translates into a meaningful operational win for the agencies that make evacuation calls.

Worth flagging that these are DeepMind's own numbers, framed against 'prior models' the blog post does not fully specify, and open sourcing weights is not the same as shipping the operational forecast pipeline national agencies actually run. The post is also quiet on how WeatherNext performs in basins outside the featured case, how often operational teams could realistically refresh a 1,000-member ensemble at scale, and how it stacks up against ECMWF's operational system on independent, full-season statistics. Still, weights, code and a free Colab is the kind of release that national weather services in cyclone-exposed regions, academic groups, and catastrophe modelers can actually try, and that shift in access is arguably the more important story here than any single benchmark number.