Overview
- Google announced WeatherNext 3 on Thursday and has begun folding its outputs into Search, Maps, Gemini and Google Cloud for researchers and developers to access.
- The model ingests low‑latency geostationary satellite mosaics and station observations to generate hourly global forecasts with station‑calibrated surface variables up to about 5 km resolution.
- Google reports large probabilistic gains for precipitation, citing up to 50% better day‑ahead accuracy and CRPS improvements versus established satellite and radar baselines.
- WeatherNext 3 produces 64‑member ensembles, direct 100‑meter wind and solar radiation estimates for renewables, and will be available via BigQuery, Earth Engine and Cloud Storage, though operational users are advised to validate it alongside existing agency systems.
- The release follows a multi‑year DeepMind research lineage and earned top placement on Brightband’s Operational WeatherBench while an arXiv paper published Sept. 4 describes the model architecture and evaluation results.