Google DeepMind says its WeatherNext AI model now forecasts where a cyclone will travel, how strong it will get, and how far its winds will reach, matching or beating the accuracy of the forecasting tools meteorologists rely on today, and the lab is releasing the model’s weights and code for free. The result, detailed in a new paper in the journal Nature, gives forecasters roughly one additional day of reliable lead time compared with the tools they used before. That figure comes from DeepMind’s own benchmarking, not an independent audit, but it lands in a field where a single day changes who gets out of a storm’s path in time.
DeepMind tested WeatherNext Cyclones against historical storms from 2023 and 2024, comparing its track predictions to the ECMWF ensemble model and its intensity predictions to NOAA’s HWRF system. Across both measures, WeatherNext’s three-day forecasts matched the accuracy that older models needed an extra day to reach. DeepMind frames the jump as equivalent to roughly a decade of typical progress in meteorology, a comparison drawn from two decades of forecast-error trend lines rather than a head-to-head test against any single successor system.
The clearest evidence of practical value arrived during last year’s Atlantic season. National Hurricane Center meteorologists used the model’s output to anticipate that a storm bearing down on Jamaica would strengthen unusually fast, and the added warning time let the agency post alerts earlier than its prior tools allowed. DeepMind built the system alongside that same Hurricane Center, CIRA (a NOAA-affiliated atmospheric research institute), and the UK’s Met Office, all of which supplied storm-tracking data and expert review during development.
The model trained on nearly 20 terabytes of atmospheric data alongside IBTrACS, a historical record spanning almost 5,000 storms, and it now runs 1,000 simulated scenarios per storm to capture rare but severe outcomes such as sudden strengthening. It does this at a resolution of 28 kilometers, roughly 100 times coarser than the specialized regional models cyclone forecasting has traditionally required. DeepMind says a full 15-day ensemble now completes in under a minute on a single TPU chip, a speed the company calls faster than scientists expected for output of this quality.
DeepMind is publishing both WeatherNext 2 and WeatherNext Cyclones on GitHub, along with a smaller WeatherNext 2-mini variant built to run inside a free Colab notebook on a single TPU. That access matters most for university labs and smaller national weather services that could not otherwise train a model at this scale on their own. It does not change who issues a hurricane warning. The National Hurricane Center, the Met Office, and equivalent regional authorities remain the only bodies with legal standing to declare evacuations, and DeepMind’s release explicitly directs users back to their local weather service for official guidance rather than positioning the model as a replacement for one.
For an industry that mostly measures AI progress in benchmark percentage points, WeatherNext is an unusual case because its central claim converts directly into hours of evacuation time ahead of landfall. Meteorological agencies testing the open weights over the coming storm season should treat DeepMind’s day-long lead-time figure as a starting benchmark to reproduce against their own regional data, not a settled result, since a self-reported comparison against two storm seasons is not the same as multi-season, independently verified forecast performance.
Published by Google DeepMind on August 6, 2026.
