What changed
Hugging Face released a public leaderboard that scores global weather models against actual weather data, showing AI models now lead in accuracy. The leaderboard includes models from national weather centers and evaluates them on daily forecasts for 24 cities across six continents, up to 10 days ahead. Darwin-Prediction-Weather, developed by VIDRAFT, has the lowest error across all four weather elements—daily maximum temperature, daily minimum temperature, daily maximum wind speed, and daily precipitation total—at every lead time from 1 to 7 days.
What this means in practice
The leaderboard provides daily updates on model performance, allowing users to see which models are most accurate for different lead times and locations. For example, Darwin-Prediction-Weather has a root-mean-square error (RMSE) of 0.84°C for daily maximum temperature at a 1-day lead time, compared to 1.35°C for the next closest model, ECMWF AIFS. The leaderboard also includes a live Earth map showing real-time weather data and the champion model for each city. Users can access the leaderboard and forecast archives through Hugging Face's platform, with direct links to specific tabs like the precipitation tab.
Limitations and availability
The leaderboard is updated daily and includes models from various weather centers, such as ECMWF, NOAA, and DWD. The Darwin-Prediction-Weather model is available for non-commercial use under the CC BY-NC 4.0 license, while the leaderboard application is released under the Apache License 2.0. Commercial use of the model requires contacting VIDRAFT. The data used for verification includes ERA5 reanalysis and Copernicus Climate Change Service, with weather data provided by Open-Meteo.com. The leaderboard and forecast archives are accessible via the provided links on Hugging Face.
