Cyclone forecasting has always forced a trade-off. A storm's track — where it's going — is steered by global atmospheric currents, best captured by large, coarse-resolution models. A storm's intensity — how strong it gets — is driven by fine-scale thermodynamics near the core, which has historically needed specialized, high-resolution local models. Running both meant stitching together two different modeling approaches and accepting the seams between them.
Google DeepMind's WeatherNext Cyclones, published August 6, 2026 in a Nature paper, collapses that trade-off into a single model — and the team is open sourcing it. This continues a lineage explainx.ai has tracked in AI's role in climate and energy sustainability, which covered DeepMind's earlier GraphCast breakthrough; WeatherNext Cyclones is the next step specifically for the most destructive weather category on the planet.
TL;DR
| Question | Direct answer |
|---|---|
| What's new? | A single model predicts cyclone track, intensity, and wind structure — no separate global/local models needed |
| How much better is it? | ~1 extra day (24 hours) of accurate lead time versus prior best-in-class models |
| Is that a big deal? | DeepMind says it's roughly a decade's worth of meteorological progress in one release |
| Has it been used in a real storm? | Yes — helped the NHC forecast Hurricane Melissa's rapid intensification in Jamaica, 2025 |
| What resolution does it need? | 28x28km — about 100x coarser than traditional intensity models require |
| Is it open source? | Yes — code, weights, and a Colab-runnable mini version (WeatherNext 2-mini) |
| Where can I see live forecasts without running it? | Weather Lab, part of Google Earth AI |
Why cyclone forecasting is worth a decade of progress in one paper
The stakes explain the urgency. Tropical cyclones — hurricanes and typhoons are regional names for the same phenomenon — have caused more than 700,000 deaths and $1.4 trillion in economic losses globally over the past 50 years, per DeepMind's announcement. Every hour of additional warning time translates directly into evacuation windows, resource staging, and lives saved.
DeepMind benchmarked WeatherNext Cyclones against historical storms from 2023-2024, comparing it to ECMWF-ENS (the operational standard for track forecasting) and HWRF (the operational standard for intensity forecasting). The result: 3-day WeatherNext forecasts match the accuracy prior systems could only deliver at 2 days out. That one-day gain compounds directly into evacuation and preparedness time for coastal communities and emergency planners.
The model already has a real-world track record. During the 2025 hurricane season, WeatherNext helped the National Hurricane Center forecast Hurricane Melissa's rapid intensification and landfall in Jamaica — documented in the NHC's 2025 Verification Report. Rapid intensification events are historically among the hardest scenarios to predict, and among the most dangerous when missed.
How WeatherNext Cyclones actually works
Three design choices explain the jump in accuracy:
1. A single model, co-trained on two data types
Instead of training separate global and local models, WeatherNext Cyclones was trained end-to-end on nearly 20 terabytes of global atmospheric data alongside the IBTrACS database — expert-curated historical records spanning almost 5,000 storms. Learning both simultaneously lets the model connect large-scale steering currents to the fine-scale processes that drive intensification, rather than treating them as separate problems handed off between two systems.
2. Functional Generative Networks for ensemble forecasting
Weather is inherently uncertain, so a single deterministic forecast understates real risk. WeatherNext uses Functional Generative Networks (FGNs) to efficiently generate large ensembles — many possible future scenarios from the same starting conditions. Last year's system produced 50 predictions per run, matching the scale of traditional physics-based ensembles. This year, DeepMind scaled that to 1,000 members per cyclone, which is large enough to surface rare but consequential tail-risk scenarios like sudden rapid-intensification events, the same category that caught forecasters off guard with Hurricane Melissa in past seasons without this tooling.
A full 15-day, 1,000-member ensemble forecast now runs in under a minute on a single TPU — fast enough for forecasters to re-run and re-evaluate probability distributions as new data comes in, rather than waiting hours for a single deterministic pass.
3. Surprisingly coarse resolution requirements
The most counterintuitive finding: WeatherNext Cyclones only needs 28x28km resolution input data — about 100x coarser than the resolution traditional intensity models have required. A smaller variant, WeatherNext 2-mini, runs at an even coarser 111x111km and still performs well. DeepMind is explicit that this surprised its own researchers, and frames understanding why coarse inputs produce this level of accuracy as an open research question for the community to help answer — a notable admission for a paper claiming a decade of progress.
What's being open sourced
DeepMind is releasing code and weights for three model variants:
| Model | What it's for | How to access it |
|---|---|---|
| WeatherNext Cyclones | The model that ran during the 2025 hurricane season; benchmarked in the Nature paper | Open weights + code |
| WeatherNext 2 | A later, operationalized update (deployed October 2025) | Open weights + code |
| WeatherNext 2-mini | Compact version, coarser resolution | Runs on a single TPU in a free public Colab notebook |
This is a meaningfully lower barrier to entry than most frontier weather models, which have historically required institutional supercomputing access. A researcher or regional meteorological agency without a dedicated compute budget can run WeatherNext 2-mini directly in a browser-based Colab session. For a broader sense of how coarse-model efficiency compares to compute-heavy alternatives, see explainx.ai's coverage of the model-selection energy math that increasingly factors into which model teams choose to run.
Anyone who doesn't want to run the model themselves can explore live predictions through Weather Lab, DeepMind's public interface, recently refreshed to show global weather variables — temperature, precipitation, wind speed — alongside cyclone tracks, all under the broader Google Earth AI umbrella.
How this fits DeepMind's weather-AI lineage
WeatherNext Cyclones is the latest entry in a sequence explainx.ai has tracked as it developed:
- GraphCast (2023) — the original breakthrough, producing 10-day global forecasts in under a minute on a single TPU v4, matching or beating operational numerical weather prediction (NWP) systems beyond day 7, as covered in explainx.ai's AI climate and energy sustainability guide.
- GenCast (2024) — added probabilistic ensemble forecasting to quantify uncertainty and extreme-weather risk, rather than producing a single deterministic forecast.
- WeatherNext Cyclones (2026) — extends the lineage specifically to tropical cyclones, unifying track and intensity prediction that previously required two separate model families, and scaling ensembles to 1,000 members.
The pattern across all three: each release trades a previously assumed hard constraint (resolution, compute time, model specialization) for a data- and architecture-driven shortcut, then open sources the result rather than keeping it proprietary. That's a meaningfully different posture than most frontier AI releases, where open weights are the exception rather than the default — worth noting for anyone tracking open-source AI momentum more broadly.
What builders and researchers can actually do with this
- Run WeatherNext 2-mini in Colab for free, no dedicated GPU/TPU cluster required — a genuinely accessible entry point for students, regional forecasters, or nonprofits building disaster-preparedness tooling.
- Explore forecasts without coding via Weather Lab, useful for product teams evaluating whether cyclone-risk data belongs in an app or dashboard.
- Read the IBTrACS + Nature paper methodology if you're building domain-specific forecasting models in adjacent areas (flooding, wildfire risk, extreme heat) — the co-training approach on global + localized expert data is broadly transferable.
- Watch the resolution question — DeepMind's admission that it doesn't fully understand why 28x28km inputs suffice is an open invitation for the research community, not a solved problem to take at face value.
What people are asking about WeatherNext
Does this replace human forecasters? No, and DeepMind is explicit about this in its announcement: the goal is a "collaborative weather forecasting ecosystem" combining machine learning with the "indispensable real-world expertise of human forecasters," not a replacement for the National Hurricane Center or regional weather services. Official warnings still come from national meteorological agencies — WeatherNext is a tool those agencies use, not a public alert system on its own.
Why 28x28km instead of finer resolution? This is the part DeepMind itself flags as unresolved. Conventional wisdom in operational meteorology held that fine-grained resolution was the main lever for accurate intensity prediction, since the physical processes driving rapid intensification happen at small spatial scales near a storm's core. WeatherNext Cyclones getting state-of-the-art results at 100x coarser resolution suggests the model is learning something about those processes indirectly, through its training data and architecture, rather than needing to simulate them directly at fine grid resolution. DeepMind is inviting the research community to help explain this rather than claiming to have already solved it.
Is this the same team behind AlphaFold and GraphCast? Yes — this is a joint effort between Google DeepMind and Google Research, continuing the same applied-science lineage that produced GraphCast (2023) and GenCast (2024), alongside DeepMind's other science-focused releases like AlphaFold and AlphaEarth. The co-author list spans both DeepMind researchers and forecasters from the National Hurricane Center, CIRA, and the UK Met Office — a genuinely cross-institutional paper rather than a purely internal DeepMind result validated after the fact.
What happens to accuracy as more agencies adopt it? That's an open operational question the paper doesn't fully answer. DeepMind's benchmarks are against historical 2023-2024 storms and one real operational season (2025's Hurricane Melissa); broader validation across more storm seasons, ocean basins, and forecaster workflows will determine how the one-day accuracy gain holds up at scale once more national weather services are relying on it as part of their standard toolkit rather than a research pilot.
Related reading
- AI's role in climate and energy sustainability — where GraphCast fits into the broader AI-for-climate picture
- The model-selection energy math — why coarser, faster models matter beyond just weather
- The AI bubble in 2026: reality check — open-source momentum as a counterpoint to bubble concerns
- Every hyperscaler nuclear deal — the compute infrastructure underneath large-scale AI-for-science work
- Google Search I/O 2026: agents and agentic coding — more on Google's 2026 AI research direction
Model benchmarks, resolution figures, and the Hurricane Melissa case study reflect Google DeepMind's Nature paper and official announcement as of August 6-7, 2026; for official storm warnings, always defer to your local meteorological agency or national weather service.