VIDRAFT Publishes Leaderboard Grading Operational Forecasts Daily

Details

VIDRAFT has released a public leaderboard on Hugging Face that grades the accuracy of major global forecast models daily, along with the multi-forecast fusion system Darwin-Prediction-Weather
In the 2026 cumulative records, the fusion engine recorded RMSE 13–39% lower than the next-best model for temperature and wind speed, and 1–4% lower for precipitation
Among meteorological agency models, AI models (ECMWF AIFS, NOAA AIGFS) showed lower errors than physics models (ECMWF IFS, NOAA GFS)
Validation on WeatherBench 2 also achieved RMSE lower than the lowest values on the official scorecard across all 15 categories, including 2m temperature, Z500, and T850
The evaluation design ensures reproducibility through common case grading, future information blocking, and delayed ground truth processing based on ERA5 reanalysis
The leaderboard app is provided under the Apache-2.0 license, while the model and data are under the CC BY-NC 4.0 license, allowing anyone to recalculate via client-side computation

Key point