The new AI weather model that updates every hour, not every six

A new AI weather model called WeatherNext 3 ingests raw satellite imagery and ground station readings, producing a neighborhood-scale forecast every hour. Traditional global models update every six hours and start from a cleaned-up "analysis" snapshot, a blended best-guess of current conditions that earlier AI models also learned from. WeatherNext 3 instead goes directly to the raw observations: geostationary satellite images (platforms parked over the equator returning imagery every few minutes) and station readings. It emits a new global forecast hourly at roughly 11-kilometer grid cells.

The paper's authors report that the model matches physics-based global models on cadence and resolution, while posting substantially lower error on near-surface temperature and dewpoint, even at stations the model never saw during training. They also describe a new state-of-the-art result for probabilistic medium-range forecasting skill, and a method that folds the normally separate data-cleanup, forecasting, and post-processing steps into a single model.

The practical case is hourly forecasts where station networks are thin: tropical cyclone tracking, solar and wind planning, and aviation routing that today waits on a six-hour cycle. What remains unknown is whether independent benchmarks confirm the error claims, and whether the model will be deployed or released to outside forecasters. The paper itself is an arXiv preprint (2609.03582) and has not been peer reviewed.