Published: 7 August 2026
Google DeepMind released the code and model weights for its WeatherNext weather forecasting models on 6 August 2026, alongside a paper in Nature. The release covers WeatherNext Cyclones, WeatherNext 2 and the compact WeatherNext 2-mini, and Google DeepMind reports that the models give forecasters more than a full day of extra lead time on cyclone track, intensity and wind structure. WeatherNext 2-mini runs on a single TPU in a free Colab notebook.
What did Google DeepMind release on 6 August 2026?
Three weathernext model sets are now public on GitHub, with both code and weights. WeatherNext Cyclones is the version that ran during the 2025 hurricane season and whose results appear in the Nature paper. WeatherNext 2 is the later update that Google operationalised in October. WeatherNext 2-mini is a compact variant that fits on a single TPU and can be run from a public Colab notebook without dedicated hardware.
The work was co-developed by Google DeepMind and Google Research with the US National Hurricane Center, the Cooperative Institute for Research in the Atmosphere and the UK Met Office. Google DeepMind frames the release as an invitation to meteorological agencies, researchers and nonprofits to build localised or specialised models on top of the published weights rather than train equivalents from scratch.
What can WeatherNext 2 and WeatherNext Cyclones do?
Cyclone forecasting has conventionally required two different modelling approaches. A storm’s track is steered by large-scale atmospheric currents, best captured by coarse global models, while its intensity is driven by fine-scale thermodynamics near the core, best captured by specialised high-resolution local models. WeatherNext predicts track, intensity and wind structure with a single model.
The models use Functional Generative Networks, an architecture for producing ensembles efficiently. Google DeepMind generates a single 15-day forecast in under a minute on one TPU. Ensemble size grew from 50 members last year, matching global physics models, to 1,000 members this year, which is what lets the system surface rare but consequential outcomes such as rapid intensification.
Resolution is the surprising part. WeatherNext Cyclones works from data at 28 by 28 kilometres, which Google DeepMind describes as 100 times coarser than traditional models, and WeatherNext 2-mini runs at 111 by 111 kilometres and still performs well. Google DeepMind states plainly that it is an open research question why the models remain this accurate at that resolution, which is an unusual admission to publish alongside a state-of-the-art claim.
WeatherNext benchmarks and how they were measured
The evaluation in the Nature paper compares WeatherNext Cyclones against leading operational models on historical cyclones. Google DeepMind reports an average gain of more than 24 hours of lead time across track, intensity and wind structure, meaning its three-day forecast is about as accurate as what prior systems delivered at two days. The company puts that step change at roughly a decade of meteorological progress based on trends over the last 20 years.
The published charts benchmark three-day track error against the ECMWF ensemble ENS and three-day intensity error against HWRF over 2023 to 2025, with WeatherNext Cyclones at roughly 100 kilometres position error and roughly 11 knots intensity error. These are Google DeepMind’s own evaluations, published in a peer-reviewed Nature paper with the National Hurricane Center, CIRA and the UK Met Office as co-authors, which is a stronger evidentiary basis than a vendor blog post but still not an independent replication.
Training used nearly 20 terabytes of global atmospheric data together with the IBTrACS historical cyclone database, covering nearly 5,000 historical storms. Co-training on both global weather dynamics and expert-curated storm observations is what Google DeepMind credits for the single-model result.
Licence, availability and hardware requirements
Google publishes the WeatherNext code under the Apache License 2.0 and the accompanying materials, including model weights, under Creative Commons Attribution 4.0. Both permit commercial use with attribution, which matters because Google DeepMind’s earlier weather weights for GraphCast and GenCast carried a non-commercial licence. For forecasting vendors, energy traders and insurers, this release is the point at which the technology becomes usable in a commercial product rather than only in research.
Hardware requirements are modest by frontier-model standards. WeatherNext 2-mini runs on a single TPU in a free public Colab notebook, and a full 15-day forecast takes under a minute on one TPU. Google DeepMind notes that WeatherNext models are not an officially supported Google product and that official forecasts and warnings should still come from a national meteorological agency.
How does WeatherNext compare to GraphCast and GenCast?
WeatherNext follows GraphCast, published in November 2023, and GenCast, published in December 2024. GraphCast established deterministic AI global forecasting; GenCast added probabilistic ensembles for extreme conditions. WeatherNext extends that line in three directions: it handles cyclone track, intensity and wind structure in one model rather than pairing a global model with a specialised local one, it scales the ensemble to 1,000 members, and its weights carry a commercial licence.
The operational track record is what separates this release from a benchmark result. During the 2025 hurricane season the model helped the National Hurricane Center forecast Hurricane Melissa’s rapid intensification and landfall in Jamaica, which allowed an earlier warning. Tropical cyclones have caused more than 700,000 deaths and $1.4 trillion in economic losses globally over the past 50 years, so lead time is the metric that matters.
The code and weights of the WeatherNext models can be found on GitHub
