By ByteBot
On August 6, Google DeepMind published a Nature paper and simultaneously released WeatherNext Cyclones on GitHub — an AI model that adds roughly one full day to tropical cyclone forecasting accuracy. This isn’t a lab benchmark. The National Hurricane Center ran WeatherNext operationally during the 2025 Atlantic season and used it to forecast Hurricane Melissa’s rapid intensification and Jamaica landfall a day before any other operational model caught it. That extra day gave emergency managers time to position resources and issue evacuation orders that would have otherwise arrived too late.
One Extra Day of Warning — Confirmed in the Field
The core claim is precise: WeatherNext’s three-day forecasts match the accuracy prior models delivered at two days. According to the DeepMind research team, this represents “roughly a decade’s worth of meteorological progress” across track, intensity, and wind structure simultaneously. Moreover, these aren’t cherry-picked metrics — the model was evaluated against ECMWF-ENS (the benchmark for track forecasting) and NOAA’s HWRF (the specialist for intensity) on storms from 2023 through 2025, across all three prediction categories.
Hurricane Melissa provided real-world proof. The storm’s rapid intensification — the kind that catches operational forecasters off-guard and drives landfall casualty spikes — was predicted by WeatherNext a day ahead. Furthermore, the NHC’s operational adoption wasn’t promotional. It followed a co-development process that included the UK Met Office, running WeatherNext alongside existing operational guidance. When it mattered on a real storm with real consequences, it delivered.
How a Coarser Model Beats the Specialists
WeatherNext Cyclones operates at 28×28 km resolution — roughly 100 times coarser than traditional high-resolution regional intensity models like HWRF. That sounds like a disadvantage. It isn’t. DeepMind trained the model on approximately 20 terabytes of atmospheric data combined with IBTrACS, a historical record of around 5,000 cyclones. Consequently, the result is a model that outperforms HWRF on intensity despite the resolution gap, suggesting that global atmospheric data carries more intensity-relevant signal than the meteorology community previously assumed.
The ensemble architecture compounds the advantage. Traditional operational runs use around 50 ensemble members to estimate forecast uncertainty. However, WeatherNext scales this to 1,000 members — generating a full 15-day forecast in under 60 seconds on one TPU. More members means better probability distributions for rare events like rapid intensification, which is precisely the failure mode behind the worst hurricane disasters. Additionally, the combination of broader training data, higher ensemble counts, and a unified model for track plus intensity plus wind structure is where the accuracy gain compounds.
Three WeatherNext Variants — Including One That Runs Free
All code and model weights are on the WeatherNext GitHub repository under Apache 2.0 (code) and Creative Commons Attribution 4.0 (weights). Both licenses allow commercial use with attribution — a deliberate upgrade from the more restrictive terms on earlier Google weather models like GraphCast. In fact, DeepMind frames this as an invitation for smaller agencies, researchers, and nonprofits to build specialized models on top of the released weights instead of training from scratch.
There are three variants. WeatherNext Cyclones and WeatherNext 2 require TPU-class hardware for production use. Meanwhile, WeatherNext 2-mini runs at 111×111 km resolution on a single TPU in a free Google Colab notebook — no specialized infrastructure required. The mini variant trades some accuracy for accessibility, but it runs on the same architectural foundation as the full model. For researchers experimenting or smaller agencies evaluating fit, Colab is the entry point.
Where WeatherNext Fits in the AI Weather Stack
This release doesn’t happen in a vacuum. NVIDIA open-sourced its Earth-2 weather models in January 2026, targeting GPU access for global weather forecasting. ECMWF has AIFS for operational prediction. As a result, WeatherNext slots in as cyclone-specialist and TPU-based, with the strongest peer-reviewed validation on track, intensity, and wind structure for tropical storms specifically. DeepMind acknowledges that AIFS wasn’t directly benchmarked against WeatherNext on cyclone intensity — a gap worth noting before making deployment decisions.
In practice, none of these models replace each other. NVIDIA Earth-2 covers general global weather at GPU scale. WeatherNext covers cyclones with higher ensemble counts and a unified model. However, ECMWF’s AIFS operates within an established operational infrastructure. Developers building climate or weather applications should therefore treat these as complementary parts of a forecast stack, not competitors where one eliminates the others.
Key Takeaways
- WeatherNext gains 24+ hours of lead time on tropical cyclone track, intensity, and wind structure — validated operationally by the National Hurricane Center on Hurricane Melissa during the 2025 Atlantic season
- The model operates at 100x coarser resolution than traditional specialists and still outperforms them, challenging assumptions about how fine-grained input data needs to be
- Three variants are on GitHub under Apache 2.0 (commercial use allowed); WeatherNext 2-mini runs free in Google Colab on a single TPU
- 1,000-member ensemble forecasts complete in under 60 seconds — better probability estimates for rare events like rapid intensification
- WeatherNext is cyclone-specific; NVIDIA Earth-2 and ECMWF AIFS cover different parts of the forecast stack and are complementary, not redundant
