WeatherNext is Google DeepMind's open repository for its global weather forecasting models, and it picked up 670 stars this week to reach 7,524 stars and 961 forks. The spike has a specific cause: release v0.3.0 on 6 August 2026 added support for WeatherNext 2, published the same day as the DeepMind blog post on cyclone forecasting and a Nature paper on operational tropical cyclone forecasting with AI. The repository itself is not new — it was created in July 2023 as google-deepmind/graphcast and later renamed, which is why the pretrained weights still live in a Google Cloud bucket called dm_graphcast.

The results are the reason to look. On cyclones, DeepMind reports that three-day forecasts are as good as what prior models provided for two days — described as more than a full 24 hours of lead-time advantage, and framed in the blog as roughly a decade's worth of meteorological progress. The blog benchmarks track against ECMWF-ENS and intensity against HWRF over 2023-2025, reporting a lead advantage on both. The model produces 1,000 possible scenarios per cyclone, and during the 2025 hurricane season DeepMind says it helped the National Hurricane Center on Hurricane Melissa, including its rapid intensification and Jamaica landfall. Partners named include the NHC, CIRA, and the UK Met Office.

WeatherNext 2 itself was announced on 17 November 2025 and is built on a Functional Generative Network (FGN), an architecture that injects noise directly into the model so forecasts stay physically realistic and internally consistent rather than averaging into mush. It covers 0 to 15 days at up to 1-hour increments, generates predictions 8x faster than the previous WeatherNext, takes under a minute on a single TPU, and DeepMind reports it surpasses the prior model on 99.9% of variables and lead times. It already powers weather in Search, Gemini, Pixel Weather, and the Google Maps Platform Weather API.

What this repository actually gives you is the runnable version of that. Weights ship for WeatherNext2_<2025 (0.25°, roughly 30 km, fine-tuned on ECMWF HRES and designed to initialise from operational HRES conditions rather than ERA5 reanalysis), plus WeatherNextCyclones checkpoints for <2025, <2024, and <2023 that reproduce the paper's per-year results, each as four model files. Crucially there is also WeatherNextCyclones Mini at 1° resolution (about 111 km) — a deliberately smaller model that forecasts the same fields, including cyclones, on a single accelerator. The Colab demo defaults to Mini on the free v5e-1 TPU runtime and walks through loading weights, running autoregressive rollouts, visualising temperature and wind, running the cyclone tracker, and taking a gradient step. The legacy GraphCast and GenCast models remain in the repo as WeatherNext Graph and WeatherNext Gen.

The limits are stated plainly by Google and deserve repeating. This is research code provided as-is, with no API stability and breaking changes possible without notice — the README recommends pinning to a release. Hardware is a real gate: the non-Mini models need an H100 for sufficient VRAM, TPU is the optimised target, and running on GPU requires switching the attention implementation. Training from scratch means pulling ERA5 and HRES data from ECMWF under their own terms. Licensing is split — Apache-2.0 for code and notebooks, CC BY 4.0 for everything else. And Google is explicit that this is not an officially supported Google product, has not been produced with or endorsed by any government meteorological agency, and does not replace official alerts or warnings.