News · 2026-08-09

Google put the hurricane model's weights on a public bucket

Google DeepMind has released WeatherNext 2 as open source with pretrained weights, including the specific cyclone checkpoint that ran live for the National Hurricane Center through the 2025 Atlantic season. The release covers model code, weight files on a public cloud bucket, and notebooks that go as far as taking a training step - not an inference-only demo. A one-degree variant is sized to run on a single GPU.

Key facts

• What is in it: model code, pretrained weights, and interactive notebooks for WeatherNext 2, WeatherNext Gen, and WeatherNext Graph; notebooks are Apache 2.0, everything else is CC BY 4.0.
• The operational checkpoint: WeatherNextCyclones_<2025, the model that ran during the 2025 Atlantic hurricane season as FNV3; the Hurricane Center's post-processed version was called GDMI.
• The speed claim: hundreds of possible weather scenarios from one starting point, each taking less than a minute on a single TPU, against hours on a physics supercomputer.
• Primary sources: Google's open-source models page and the google-deepmind/weathernext repository, which now carries 7,100 stars and 939 forks.

Weather forecasting has spent decades running physics simulations on some of the largest computers on earth: encode the equations governing the atmosphere, seed them with today's observations, and integrate forward. It works, and it is punishingly expensive, which is why forecasters have historically been able to afford only a handful of alternative futures per run.

WeatherNext 2 replaces the simulation with a learned model, and the interesting part is how it produces a spread of outcomes. Google DeepMind's own description of the payoff is concrete: "WeatherNext 2 can predict hundreds of possible weather outcomes from a single starting point. Each prediction takes less than a minute on a single TPU; it would take hours on a supercomputer using physics-based models."

The mechanism has a name - a Functional Generative Network - and a genuinely elegant trick behind it, described in "Skillful joint probabilistic weather forecasting from marginals". Training a model to get the joint picture right is hard: you need it to know that if it is unusually wet here it is probably also windy three hundred miles that way, and there is no clean training signal for that. So the model is instead trained only on marginals - one variable at a time, at one place at a time - and the coherence is created by injecting noise into the network's own function space rather than into its output. Each perturbation produces a whole physically-consistent alternative world rather than static scattered across a map.

The analogy that fits is a jury. Ask one forecaster and you get one confident answer. Ask sixty-four independently perturbed forecasters and the disagreement is the product: when they cluster, you have a confident forecast, and when they scatter, you have quantified uncertainty. WeatherNext 2 runs a 64-member ensemble at quarter-degree resolution over a fifteen-day horizon, which is what makes worst-case planning possible - and worst cases are the ones that matter.

Why the weights release matters more than the usual open-source announcement: this is not a research artifact that was never used in anger. The National Hurricane Center partnered with Google to run this model live for forecasters, and the released repository names the exact checkpoint. When the system called Hurricane Melissa's landfall five days out, that was this lineage of model. Anyone can now download the same weights, reproduce the paper's results on 2023, 2024, and 2025 seasons, and check the numbers themselves - which is a materially different proposition from reading a Nature paper about a system you can only see through a vendor's API. The cyclone work is published as "Operational Tropical Cyclone Forecasting with AI", with authors drawn from both DeepMind and the Hurricane Center itself.

There is also a deliberate on-ramp. WeatherNext Cyclones Mini is a one-degree version the repository describes as "a lightweight version suitable for lower memory and compute constraints (e.g., local testing or single TPUs or GPUs)," forecasting the same quantities including cyclones. It is coarser, and Google says plainly it is "not expected to match the performance of the larger versions." But it means a graduate student or a national met service without a supercomputer budget can put hands on the actual architecture rather than a description of it.

The honest caveats are worth stating. The licensing is split - the notebooks and code are Apache 2.0, but everything else including the weights falls under CC BY 4.0, which is a content license rather than a software one and carries attribution obligations that a team planning a commercial product should read carefully. This is open weights in the useful sense, not a blanket permissive grant. And a forecast model is only as good as what you feed it: the operational checkpoints are fine-tuned to be initialized from a European weather agency's operational analysis rather than from reanalysis data, so running them well means having access to that upstream feed. Downloading the weights is the easy half.

Primary source, verified: read the paper → (arXiv 2506.10772)

Ground Truth. (2026, August 9). Google put the hurricane model's weights on a public bucket. Ground Truth. https://groundtruth.day/news/google-put-the-hurricane-models-weights-on-a-public-bucket.html

@misc{groundtruth:google-put-the-hurricane-models-weights-on-a-public-bucket,
  title  = {Google put the hurricane model's weights on a public bucket},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/google-put-the-hurricane-models-weights-on-a-public-bucket.html}
}

Topics: open-weights · google-deepmind · weather · science · forecasting · ensembles