Korean startup VIDRAFT's Darwin-Prediction-Weather AI model outperformed eight major forecasting systems — including ECMWF AIFS, NOAA AIGFS, ECMWF IFS, NOAA GFS, DWD ICON, JMA GSM, Environment Canada GEM, and MET Norway MET Nordic — across all 28 evaluated categories (four variables × seven lead times) in a nine-month, 24-city study spanning February 1 to November 11, 2026, covering 1.23 million forecast-observation pairs. For 7-day maximum temperature across 4,896 matched cases, Darwin achieved an RMSE of 2.08°C, 14.0% lower than ECMWF AIFS (2.42°C), 36.1% lower than NOAA AIGFS (2.83°C), and 37.5% lower than ECMWF IFS (2.86°C). At 1-day lead, Darwin's max temperature RMSE of 0.84°C was 38.6% smaller than AIFS's 1.36°C. For 7-day minimum temperature, Darwin reached 1.81°C versus AIFS's 2.18°C (16.7% improvement), and for 7-day wind speed, 4.36 km/h versus AIFS's 5.64 km/h (22.9% improvement). Darwin's 7-day max temperature RMSE (2.08°C) undercut AIFS's 6-day RMSE (2.14°C) and IFS's 5-day RMSE (2.21°C), implying a one-day and two-day effective horizon advantage, respectively. In precipitation binary detection (≥1mm), Darwin scored 90.6% next-day and 79.6% at 7-day versus AIFS's 88.4% and 77.2%. Darwin ranked first in 7-day wind speed for all 24 cities, minimum temperature for 20 cities, and maximum temperature for 18 cities, maintaining top rank across rolling 7-, 14-, 30-, and 90-day windows. Results were submitted to WeatherBench2 for standardized comparison. No model architecture, training data, or public access has been disclosed.
Read original
dev.to