Google DeepMind's WeatherNext 3, a weather forecasting model, uses station data for hourly 5 km forecasts.

Editorial illustration for Google DeepMind's WeatherNext 3 Uses Station Data for Hourly 5 km Forecasts

Google DeepMind's WeatherNext 3 Uses Station Data for...

By AI Daily Post Edited by Brian Petersen, Editor-in-Chief

September 5, 2026 • 3 min read

Google DeepMind put a new number on the board Tuesday: 5 kilometers, updated every hour, anywhere on Earth. WeatherNext 3, built by Google DeepMind and Google Research, is the latest attempt to close two gaps that have dogged AI weather models since they started beating physics-based forecasts three years ago. The first is resolution.

Most global AI models output grids too coarse to capture what a mountain range or a coastline does to local weather. The second is timing. Standard numerical weather prediction analysis, the data most AI models train and initialize on, lands roughly six hours after the fact.

WeatherNext 3 goes after both problems at once. It reads a live geostationary satellite mosaic straight into the model, re-initializes hourly instead of waiting on stale analysis, and pushes output down to 0.05 degrees, about 5 km. It also trains against raw weather station readings, not just reanalysis grids, which matters more than it sounds like on paper.

Brightband's independent live evaluations, according to Google AI, now rank it the most accurate global weather model available. Whether outside groups get to use it is a separate question entirely.

WeatherNext 3, released by Google DeepMind and Google Research, attacks both. It takes a live global geostationary satellite mosaic as a direct model input, re-initializes every hour, and emits forecasts down to 0.05° (~5 km) while training against raw weather station measurements rather than reanalysis grids alone.

Google DeepMind’s WeatherNext 3 Trains on Weather Station Observations to Deliver 5 km Global Forecasts, Refreshed Every Hour - MarkTechPost

Why this matters

For anyone building on top of weather data, WeatherNext 3's hourly refresh cycle matters more than the headline resolution number. Six-hour-old NWP initialization has been the quiet bottleneck for every AI forecaster since this race started three years ago, and Google DeepMind just removed it by feeding in live geostationary satellite mosaics instead of waiting on analysis products. That's a real architectural shift, not a tuning improvement.

We'd flag one thing worth watching: training still leans on ERA5 and HRES-fc0, the same reanalysis sources that smooth out the coastline and valley detail this model claims to resolve at 5 km. Mixing station observations and IMERG rainfall data into training helps, but the proof is in how well 0.05° outputs hold up against actual terrain-driven weather, not benchmark scores against coarser models. For founders building downstream products, faster refresh cycles open real opportunities in logistics, agriculture, and disaster response, provided the local accuracy claims survive contact with messy, non-idealized terrain.