Introducing WeatherNext 3, our most advanced and accurate global weather AI model

What it claims

WeatherNext 3 replaces WeatherNext 2, which forecast on a 25 km grid in 6-hour steps, with hourly global forecasts at three resolutions:

• 5 km for key surface variables (temperature, moisture).
• 10 km for other surface variables.
• 25 km for atmospheric variables such as wind speed.

Google says the result is roughly five times sharper than WeatherNext 2. The main change is the training data. WeatherNext 2 learned from numerical weather prediction (NWP) output, which arrives about 6 hours late. WeatherNext 3 learns from live observations instead, and the authors say that lag causes bias in fast-changing variables like rain and surface temperature.

How it works

flowchart LR
  A[Live 1-h geostationary satellite mosaics] --> M[Single FGN mesh transformer]
  B[Historical analysis] --> M
  C[Sparse station observations] --> M
  D[IMERG + Google satellite-radar precip reanalysis] --> M
  M --> G[Dense gridded fields]
  M --> T[Cyclone tracks]
  M --> S[Station-level point forecasts]

• Inputs: live 1-hour geostationary satellite mosaics, alongside traditional historical analysis.
• Model: a single Functional Generative Network (FGN) mesh transformer.
• Station data: the model trains directly on sparse weather-station observations. This is what lets a global 5 km grid reflect local topography.
• Precipitation targets: NASA's IMERG and Google's own global precipitation reanalysis built from satellite radar.
• Outputs: one model natively produces gridded fields, discrete cyclone tracks, and forecasts at station coordinates.
• Energy variables: 100 m (turbine-height) wind, cloud cover and solar radiation.

Evidence

• Medium-range precipitation CRPS: improves by up to 60% against IMERG, 30% against MRMS, and 10% against rain gauges, at early lead times. The page says "against baselines" but does not name them.
• Consumer products: precipitation forecasts a day or more ahead are up to 50% more accurate. The page says the largest gains are in regions where forecasts were historically less reliable.
• Ranking: the "most accurate" claim rests on Brightband's independent live leaderboards. The page gives no scores from them.
• Figures:
• Figure 2 compares 2 m temperature over the UK at 0.25° (WeatherNext 2) and 0.05° (WeatherNext 3).
• Figure 3 compares probability of precipitation above 1 mm. Note that the WeatherNext 3 panel there is at 11 km, not 5 km.

How to use it

• No model setup: the page describes ready-made hourly forecast data rather than a model you run.
• Data access: query it in BigQuery or Earth Engine, or bulk-download it from Google Cloud Storage.
• Product surfaces: the forecasts also feed Google Search, Gemini, Maps, the Google Maps Platform Weather API and Earth Engine.
• Not stated: the page mentions no model weights, licence, pricing or API schema. It links to a paper and a "Build with WeatherNext 3" page, which are the places to check.

Limits

• Unknown comparison: the CRPS baselines are unnamed, and the gains are stated as "up to" and limited to early lead times. You cannot tell from this page what WeatherNext 3 was compared against, or how it does at longer lead times.
• Resolution varies by variable: only some surface fields reach 5 km. Atmospheric fields, including wind speed, stay at 25 km, so check the actual grid for any variable you plan to use.
• No official-warning use: the page says it is not a replacement for official forecasts, severe-weather warnings or safety advisories.
• Blog only: the details of the method, its compute cost, and the evaluation are in the linked paper, not on this page.