DeepMind Open-Sources WeatherNext Cyclones After a Nature-Verified Breakthrough

Google DeepMind published a Nature paper on August 6 showing its WeatherNext model achieves state-of-the-art accuracy on tropical cyclone track, intensity, and wind structure - and then did the rare thing: open-sourced the model weights and code on the same day. The WeatherNext repo is live with 6.9k stars, Apache-2.0 code, and the checkpoint files that ran live during the 2025 Atlantic hurricane season.

The headline number: on average, WeatherNext's three-day forecasts are as accurate as what prior operational models delivered in two days. That is an extra 24 hours of lead time, which DeepMind says corresponds roughly to a decade of meteorological progress on cyclone prediction. Tropical cyclones killed more than 700,000 people and caused $1.4 trillion in economic losses over the past 50 years, so this is a category where one day of warning is measurable in lives.

What the Paper Actually Shows#

The work, "Operational Tropical Cyclone Forecasting with AI", is a collaboration between DeepMind and Google Research, the US National Hurricane Center (NHC), CIRA, and the UK Met Office. It is not a lab demo: the model ran operationally during the 2025 hurricane season and helped the NHC issue what DeepMind calls a historic forecast for Hurricane Melissa - predicting rapid intensification and landfall in Jamaica early enough for ground teams to prepare. The NHC's 2025 verification report covers the operational results.

The technical story has three pieces:

One model, two regimes. Cyclone track (where the storm goes) is steered by huge global atmospheric currents, while intensity (how strong it gets) is driven by fine-scale thermodynamics around the core. Historically those demanded two different model families: coarse global models for track, high-resolution local models for intensity. WeatherNext is a single model that predicts track, intensity, and wind structure together, trained end-to-end on nearly 20 terabytes of global atmospheric data plus the IBTrACS database of roughly 5,000 historical storms.

Low resolution, high accuracy. The model operates at 28x28 km input resolution - 100x coarser than traditional operational models - and beats them anyway. That result "surprised scientists," per the post, and remains an open research question. A mini variant at 111x111 km also performs well.

Ensembles from Functional Generative Networks. The uncertainty quantification comes from FGNs, described in the technical report: ensembles generated via learned model perturbations, trained directly on CRPS (a proper scoring rule for probabilistic forecasts). This year the system runs 1,000-member ensembles per cyclone - up from 50 last year - to capture rare tail events like rapid intensification. A single 15-day forecast takes less than a minute on a TPU.

What's Open Now#

The release covers three model families in one repo:

• WeatherNext 2, the global medium-range atmospheric model, including the operational checkpoint initialized from ECMWF HRES data
• WeatherNext Cyclones, the operational cyclone model with three checkpoints (trained through 2022, 2023, and 2024) that reproduce the paper's results
• WeatherNext 2-mini / Cyclones Mini, a 1-degree lightweight version that runs on a single TPU or a P100 GPU

The mini model is the developer-friendly entry point: the repo ships a free Colab notebook on the v5e-1 runtime that loads weights, runs autoregressive rollouts, and even runs the direct cyclone tracker. The non-mini checkpoints need an H100 on GPU or a v5p TPU.

There is also a data-access path for people who do not want to run the model at all: daily WeatherNext forecast outputs are published through Google Cloud (Earth Engine, BigQuery, and Vertex AI), the Weather Lab visualizer, and an OpenMeteo API. The repo also carries the older GraphCast and GenCast code, so the full lineage is in one place.

What Developers Are Saying#

The discussion around the release split along three lines, and the skeptics made fair points worth carrying into your own evaluation.

The dominant mood was appreciation that a frontier lab shipped a problem-specific model instead of another general agent. Several commenters called weather the rare domain where ML surrogates actually beat physics-based models, and pointed out that the underlying architectures - graph neural networks, and now functional generative networks - get far less attention than they deserve.

The skepticism clustered on the "extra day" framing. The honest read of the paper's own charts: WeatherNext's 3-day error equals the prior models' 2-day error, but that is not the same as proving the extra day changes evacuation decisions. Commenters working in logistics and emergency planning pushed back, noting that evacuating hospitals, prisons, and elderly populations takes days, and that a more confident earlier warning changes how much expensive equipment and shipping can be moved out of surge zones. The counterpoint was that forecast value depends on confidence, not just lead time - a vague "something might hit somewhere" does not trigger costly action, and that is exactly what ensembles with 1,000 members are for.

The third thread was about the data foundation. ML weather models are trained on ERA5 reanalysis - itself a physics-based product - and initialized from operational NWP fields, so the public infrastructure of weather observation (balloons, satellites, stations) is doing silent heavy lifting behind every "AI beats physics" headline. One commenter noted ECMWF's own AI ensemble has been operational since mid-2025, so DeepMind is not alone in production AI weather; and a PyTorch reproduction via NVIDIA's PhysicsNeMo project appeared within hours of the release.

Why This Matters for Developers#

Three takeaways land for anyone building with or around AI:

1. The model-to-data release pattern is the template. Weights and code are useful, but the forecast data feeds - BigQuery, Earth Engine, OpenMeteo - are what let a developer build a weather-aware product this week without touching a TPU. That mirrors the shift we covered in open-weights economics: the open release is only the start; the hosted inference and data products are where the value compounds.

2. CRPS-trained ensembles are a pattern worth stealing. Training an ensemble that is jointly consistent with a proper scoring rule, not just individually accurate, is exactly the framing that makes probabilistic forecasts actionable. It is the same lesson as our leanstral coverage: narrow, well-scoped models with honest uncertainty beats broad models with vibes.

3. Verification beats press releases. The strongest evidence in the announcement is not the Nature figure - it is that the NHC ran the model operationally, published a verification report, and the 2025 season produced a forecast (Melissa) that the agency credited. For a field drowning in vendor-reported benchmarks, an operational trial with a government partner is the difference between a claim and a deployment. Compare that bar with how we grade AI benchmark claims generally.

If you want to poke at the model directly, the Colab notebook is the fastest path - you can have a 15-day forecast for a real cyclone in the time it takes to read the README. And if you want the bigger picture on where DeepMind sits in the open-weights landscape, Gemma 4's release and the open-weights leadership debate are the surrounding context.

Continue Reading#

• Gemma 4: DeepMind's Open-Weight Answer
• Leanstral 1.5: A Theorem-Proving Model Built for One Job
• Self-Hosting Open-Weight Models: The Break-Even Math
• Your Benchmark Is Lying to You
• Gemini Robotics 2: Whole-Body Intelligence in Production

Sources#

• DeepMind blog: WeatherNext breakthrough in forecasting cyclones (August 6, 2026)
• Nature paper: Operational Tropical Cyclone Forecasting with AI
• GitHub: google-deepmind/weathernext (weights, code, Colab notebook)
• FGN technical report on arXiv
• NHC 2025 Verification Report
• Weather Lab and OpenMeteo WeatherNext API