WeatherNext: Google's Cyclone AI Adds a Day of Lead Time

Google DeepMind published a cyclone-forecasting AI in Nature on August 6, reporting that WeatherNext Cyclones predicts a storm’s track, intensity and wind structure with a day or more of extra lead time over leading operational models — its three-day forecasts matching what two-day forecasts used to deliver. The model weights are now free to download, the US National Hurricane Center already used it through the 2025 season, and for cyclone-prone countries like India the question shifts from “does AI forecasting work?” to “who puts it to work first?”

What DeepMind actually announced

The announcement bundles three things: a peer-reviewed result, an operational track record, and an open-source release.

The result is the paper, “Operational Tropical Cyclone Forecasting with AI”, which evaluates WeatherNext Cyclones on tropical cyclones from 2023 to 2025 and finds its track, intensity and wind-radii predictions “offer an average of a day or more of lead time advantage over leading operational models.” DeepMind’s own announcement puts the same finding more plainly: “our three-day forecasts are as good as what prior models were able to provide for only the next two days,” a jump the company frames as “roughly a decade’s worth of meteorological progress,” measured against 20-year improvement trends in the benchmark systems — ECMWF’s ensemble for track, and the US HWRF model for intensity.

The operational record is the 2025 hurricane season. The model ran live with the US National Hurricane Center (NHC), and DeepMind says it helped forecasters anticipate Hurricane Melissa’s rapid intensification before its Category 5 landfall in Jamaica — the company’s May 2026 case study says the model signalled a Jamaica landfall five days out at 80% confidence. This year the system scaled from 50 simultaneous scenarios to a 1,000-member ensemble, the better to catch low-probability, high-consequence turns like rapid intensification.

The release is the part developers and weather agencies will feel. The code and model weights are on GitHub, a compact WeatherNext 2-mini runs on a single TPU in a free Colab notebook, and forecasts are browsable in DeepMind’s Weather Lab. This follows the open-weights turn we have tracked across the industry — from Alibaba’s Qwen3.8-Max to the broader surge of open Chinese models — but weather may be the domain where free weights matter most, because the users who need them (national met agencies, disaster authorities, researchers) are exactly the ones who cannot pay frontier-lab prices.

Two details from the paper’s own framing deserve attention. First, the model was trained on nearly 20 terabytes of atmospheric data plus the IBTrACS archive of roughly 5,000 historical storms, and generates a full 15-day forecast in under a minute on a single TPU — physics-based models need supercomputer hours. Second, it works from input data at 28×28 km resolution, about 100 times coarser than the specialised models meteorologists believed intensity forecasting required. DeepMind’s post says plainly that this “has surprised scientists, and it remains an open research question to fully understand how our models produce such accurate predictions at this resolution.” A breakthrough the builders cannot yet fully explain is still a caveat, whatever the benchmark says.

How the claim measures against today’s best forecasts

DeepMind published a relative claim — an extra day of lead time — rather than a headline error figure, so the honest way to read it is against what official agencies already achieve. Their own verification reports are the yardstick: the NHC’s 2024 report calls that Atlantic season its most accurate on record, and IMD’s RSMC New Delhi 2024 report documents India’s five-year averages.

Sources: NHC Verification Report 2024; RSMC New Delhi Annual Verification 2024. Nautical miles converted at 1.852 km. The two basins breed different storms in different data environments, so the columns are context, not a contest.

Against curves like these, “three-day forecasts as good as two-day forecasts used to be” is an enormous claim — at NHC’s 2024 skill, it would be like pulling the 124 km error at 72 hours down toward the 84 km error at 48 hours. What the claim is not, yet, is independently re-verified: coverage of the announcement — Engadget, Tech Times and others — relays DeepMind’s numbers rather than auditing them. Peer review in Nature and a season inside NHC’s workflow are real validation, but the only outside meteorologist verdict on record dates to the model’s experimental phase, when former NHC hurricane-specialist chief James Franklin told CBS News Miami during 2025’s Hurricane Erin that “the Google DeepMind model did it a little bit better than any of the other ones” on track — while cautioning that no hurricane model is perfect.

The agencies themselves are measured. The US National Weather Service’s June 2026 statement on adopting AI hurricane models — DeepMind’s included — reads: “Now we are ready to begin incorporating the new AI models into our toolbox, although the learning will continue!” Guidance in the toolbox, not the official forecast. DeepMind’s own post carries the same disclaimer: for warnings, refer to your national weather service.

The India angle

India’s cyclone problem is the Bay of Bengal, and its progress is real but uneven. IMD’s average track errors for 2020–24 stood at 72, 111 and 154 km at 24, 48 and 72 hours — better than 81/126/171 km in 2015–19, and the Ministry of Earth Sciences told the Rajya Sabha in March 2026 that track accuracy improved 20–25% over five years, with landfall and intensity accuracy up 35–45%. But the line does not only go up: in 2025 the track errors ran worse than the five-year average — 80, 120 and 204 km — even as intensity errors were sharply better than average, a reminder that a hard season can still bend the curve.

What those kilometres buy is measured in evacuations. The 1999 Odisha Super Cyclone killed roughly 10,000 people; Cyclone Phailin in 2013, of comparable strength, killed 23 — the Odisha government’s own review credits planning and evacuation, both children of forecast lead time. The pattern held through recent seasons: Cyclone Dana in October 2024 saw around a million people moved and no direct deaths in Odisha, while Cyclone Ditwah in late 2025 killed over 600 in Sri Lanka against three in Tamil Nadu. Every extra day of credible warning is more districts emptied in time.

Here is the catch: India does not yet run a dedicated AI model for cyclones. The March 2026 parliamentary answer describes IMD’s cyclone stack as numerical weather prediction, multi-model ensembles and a decision-support system. India’s AI deployments are adjacent: two AI-enabled monsoon tools went live in May 2026 — block-level onset forecasts and a 1-km nowcasting pilot — and the ₹2,000-crore Mission Mausam explicitly plans to fuse AI/ML with physics models through 2031. The Bharat Forecast System launched in May 2025 sharpened the physics grid from 12 km to 6 km, but it is a numerical model, not an AI one.

That gap is exactly what free weights address. A model that runs on a single TPU — with a mini version in a free Colab — is testable by IITM Pune or IMD’s research wing this season, without procurement, licensing or a data-sharing agreement. There is precedent for Google weather tech operating at Indian scale, but note who signed it: Flood Hub grew from a 2018 Patna pilot to cover about 200 million people in India in partnership with the Central Water Commission — the water ministry’s agency, not IMD. Neither IMD nor the Ministry of Earth Sciences had commented publicly on WeatherNext as of August 9; whether the Bay of Bengal’s gatekeeper engages is the India question this release poses.

What to watch

The October–December test. IMD’s cyclone page showed no active system in Indian waters as of August 8, and Bay of Bengal cyclogenesis peaks post-monsoon. If WeatherNext output starts appearing in Indian research commentary during the coming season, adoption has begun; silence will be informative too.

Whether anyone independently reproduces the numbers. The weights being public means ECMWF, academics — or IITM — can now re-run DeepMind’s verification instead of taking Nature’s word for it. Worth recalling that ECMWF stopped hosting external AI models in May 2026 after an upgrade of its own system degraded their tuned performance — a sign of how fast this field’s pecking order shifts. ECMWF’s own AIFS has been operational since February 2025 claiming up to 20% better tropical cyclone tracks, and NOAA switched on its own AI global models in January 2026 — DeepMind is ahead on cyclone specifics, not alone.

The repo catching up to the announcement. When we checked, the GitHub repository’s visible documentation still centred on the older GraphCast and GenCast checkpoints; the cyclone-model artefacts were not yet clearly labelled. Announcement-to-repo lag is common, but “open” only counts once the cyclone weights are verifiably downloadable.

How agencies phrase attribution. The NHC has begun working AI models into its forecast process this season, per its June 2026 statement — and every mention in official products normalises the technology for other agencies, IMD included. The rest of our AI models coverage has mostly tracked labs racing each other on benchmarks, like DeepSeek’s quiet price-performance climb; this is the rarer story where the benchmark is a coastline.

For official storm forecasts and warnings, rely on your national meteorological agency — IMD (mausam.imd.gov.in) in India, the NHC in the US.