One-line takeaway

Google DeepMind and Google Research published WeatherNext Cyclones (WN-C) in Nature, open-sourcing the model and weights. According to the same paper, this AI weather model uses inputs roughly 100× coarser than traditional regional models and still adds an average of more than one day of lead time on tropical cyclone track, intensity, and wind-structure forecasts.

Background: why hurricanes are hard to forecast

Tropical cyclones are among the most dangerous and costly weather phenomena. The DeepMind launch blog cites more than 700,000 deaths and $1.4 trillion of economic losses globally over the past 50 years. Hurricane forecasting has long been split between two regimes: tracks depend on the global atmospheric circulation, where coarse global models excel; intensity is driven by highly localized thermodynamic processes around the cyclone core, where higher-resolution regional models excel. The two have to be stitched together by human forecasters.

The WeatherNext Cyclones approach

WN-C is a single end-to-end AI model that outputs track, intensity, and wind structure jointly. Key design points (per the Nature paper, the DeepMind launch blog, and the open-source repository):

  • Training data: ~20 TB of global atmospheric reanalysis (built on ECMWF) plus the IBTrACS historical cyclone database of nearly 5,000 storms.
  • Architecture: built on Functional Generative Networks (FGN, arXiv:2506.10772) for ensemble generation. A single 15-day forecast finishes in under a minute on a TPU, with up to 1,000 physically consistent members (vs. the conventional 50).
  • Input resolution: only 0.25° (28 km), roughly 100× coarser than traditional regional models. A 1° (111 km) Mini version is also released, runnable on a single TPU through Colab.
  • Counter-intuitive result: the authors explicitly call out that achieving state-of-the-art intensity forecasts from such coarse inputs is itself an open research question — they are not fully sure what signal the low-resolution inputs encode.

Empirical performance (Nature + Ars Technica)

The Nature paper's abstract headline number: across 2023–2025 retrospective evaluation, WN-C delivers an average lead-time advantage of more than a full day for track, intensity, and wind radii, "comparable to the progress seen over the last decade of operational development."

Finer-grained numbers from the DeepMind launch blog (2026-08-06) and Ars Technica (2026-08-08):

  • 2025 live deployment — the U.S. National Hurricane Center used WeatherNext operationally and, for the first time, predicted a Category 5 landfall in Jamaica while Hurricane Melissa was still a Category 1. The blog quotes a 5-day, 80% confidence prediction that the system would hit Jamaica as a Category 5.
  • Ensemble scale-up — single-storm ensembles went from 50 members in 2024 to 1,000 in 2025, with the goal of capturing rare rapid-intensification events.
  • Open-sourcing — code, weights, WN-C, and WeatherNext 2 are all released together in the google-deepmind/weathernext GitHub repository, with operational data feeds pushed through Google Cloud (Vertex AI / BigQuery / Earth Engine), the Weather Lab visualization page, and Open-Meteo.

Where this sits relative to prior AI weather models

WeatherNext itself is not new. DeepMind's GraphCast (2023) and GenCast (December 2024) already pushed AI weather forecasting into ECMWF-class territory. The change here is that WeatherNext Cyclones is the first to complete a full loop: Nature peer review + NHC operational integration in 2025 + public release. It also locks in a counter-intuitive empirical conclusion — high resolution is not a strict prerequisite for state-of-the-art intensity forecasting — which begins to displace a long-standing default assumption.

Calibration of the key numbers

  • Input resolution: 0.25° (~28 km). Confirmed by the Nature paper, DeepMind blog, and GitHub README; the ~100× coarseness figure is restated by both the DeepMind blog and Ars Technica.
  • Ensemble size: 1,000 members in 2025 (50 in prior years), per the DeepMind blog.
  • Historical cyclone training data: ~5,000 storms in IBTrACS, per the DeepMind blog.
  • Operational milestone: first NHC operational use of WeatherNext was in the 2025 Atlantic hurricane season, with the 5-day, 80% Category 5 prediction for Jamaica attributed to NHC director Mike Brennan in Ars Technica's 2026-08-08 report.
  • Paper publication: 2026-08-06, DOI 10.1038/s41586-026-10953-2 (Nature).

What this means in practice

For AI researchers: the "high resolution = better" default is now loosened by peer-reviewed evidence. Compute budgets for weather AI training can be reallocated; running a 1° model first and only scaling to 0.25° after proof-of-concept is now a cheaper research path. The Mini release lowers the entry cost to a single TPU on Colab.

For national weather services: the path used by NHC in 2025 is now open-sourced. Other meteorological agencies can fork the same architecture and training recipe rather than starting from scratch.

For readers watching general AI progress: this is the latest example of a foundational research result rewriting a mission-critical operational pipeline. It belongs to the same era signal as Microsoft's shift away from tokenmaxxing and Oracle's OpenJDK ban on AI-generated code — AI is simultaneously a tool and a new dependency, and its failure modes need to be audited earlier than its adoption curve.

Primary references