Darwin-Prediction-Weather tops every lead time on 2026 weather board

Forecasting has long relied on physics-based models that solve atmospheric equations on supercomputers, while AI models trained on historical weather data have joined national agencies over the past three years. The leaderboard addresses the practical question of which to trust for tomorrow, three days out, or a week ahead by scoring every model on the same cases: only days and cities where all models issued a forecast. VIDRAFT's Darwin approach combines multiple strong models with skill-based weights, adds location-specific corrections for temperature and wind, and trains only on observations confirmed before the forecast being scored, so no future information enters. Beyond the 2026 board, Darwin was evaluated on WeatherBench 2, Google Research's standard benchmark for AI weather models, and recorded the lowest error in all 15 evaluated cells under the 2020 deterministic protocol, covering 2 m temperature, 500 hPa geopotential, and 850 hPa temperature at 1-, 3-, 5-, 7-, and 10-day lead times. The forecast files were published and submitted to the WeatherBench 2 team.