Google Launches WeatherNext 3, Its Most Advanced Global Weather AI Model

On September 3, Google DeepMind and Google Research launched WeatherNext 3, calling it a major breakthrough in global weather forecasting and their most advanced and accurate global weather AI model to date. It learns directly from real-world, real-time observations rather than traditional numerical methods, and tops Brightband's real-time leaderboard. Google also opened access to building on WeatherNext 3 and announced the Google Earth AI planetary intelligence initiative.

Confirmed

• Hourly refresh: The model learns directly from geostationary satellite data, bypassing the roughly 6-hour data lag of traditional numerical forecasts, producing new high-resolution forecasts every hour at up to 5 km resolution—valuable for fast-developing severe weather warnings.
• Performance gains: Precipitation forecast error drops by up to 50%, with the largest improvements where historical forecasts were unreliable; team-supplied details cite CRPS improvements of up to 60% over IMERG, and forecast detail up to 5x finer than WeatherNext 2, capturing rapidly evolving storms, local temperature changes, and clean-energy-relevant variables.
• Data sources: Training incorporates new real-time satellite data; using raw satellite data is the key differentiator from prior models.
• Product integration: WeatherNext 3 will power weather in Google Search, the Gemini app, Google Maps, and Google Maps Platform; developers and researchers can access it via BigQuery. DeepMind research lead Pushmeet Kohli reshared the announcement, noting the team is moving fast in this direction.

Why it matters

• Learning directly from observations with hourly refresh and up to 5 km resolution makes forecasts more localized, precise, and timely.
• Up to 50% lower precipitation error and up to 60% CRPS improvement directly benefit high-impact weather warnings and public safety.
• Open access via Search, Gemini, Maps, and BigQuery, plus the Google Earth AI initiative, brings the results quickly to users, developers, and researchers.

2026-09-03 ~ 2026-09-04 · 19 related posts

• [source] DeepMind launches WeatherNext 3: hourly-refresh AI weather forecasts at 5km resolution — Google DeepMind · 2026-09-03
• WeatherNext 3 now powers forecasts in Google Search, Gemini, and Maps, with data via BigQuery — GoogleDeepMind · 2026-09-03
• Google launches WeatherNext 3, its most advanced global weather AI model — Recoil42 · 2026-09-04
• [source] WeatherNext 3 details: real-time satellite training, hourly 5-km forecasts, 60% CRPS gain vs IMERG — ymatias · 2026-09-04

15 near-duplicate retellings: PeterWBattaglia · GoogleDeepMind · GoogleDeepMind · ZoubinGhahrama1 · GoogleDeepMind · Scobleizer · geoffwolfe · GoogleAI · GoogleAI · pushmeet · rseroter · ymatias · ymatias · anselm · aigclink