DeepMind open-sources WeatherNext cyclone-forecast weights
Open weights, a free Colab, and Earth Engine + Vertex access — plus the non-commercial license catch. DeepSeek's V4 Flash 0731 also tops the boards.
Copy markdown
Open weights for a frontier weather model
DeepMind open-sourced WeatherNext 2 and WeatherNext Cyclones — the GraphCast and GenCast weights it ran through this hurricane season. The code is Apache 2.0 on GitHub, and a WeatherNext 2-mini notebook runs free in Colab, so you can generate forecasts without a supercomputer.
A decade of forecasting, in one release
Its three-day track forecasts are as accurate as older models' two-day ones — about 100 km position error and 11-knot intensity error at 72 hours. DeepMind frames the leap as roughly a decade of meteorological progress.
Build a weather app this weekend
A full 15-day forecast renders in under a minute on a single TPU, and the model now emits 1,000-member ensembles (up from 50 last year). Pull gridded outputs straight from the Earth Engine catalog or run inference on Vertex to wire probabilistic weather into a product.
Check the license before you ship
The GitHub weights are CC BY-NC-SA 4.0 — non-commercial only, even though the code itself is Apache 2.0. For anything paid, route through Vertex, the Earth Engine dataset, or Google's Weather API instead of the raw checkpoints.
Elsewhere: DeepSeek's V4 Flash 0731 tops the cheap tier
A new DeepSeek V4 Flash 0731 checkpoint scores 50 on the Artificial Analysis Intelligence Index — up 10 points from the prior build — and has its ARC-AGI numbers posted on ARC Prize's official board, all at V4 Flash's bargain token price.