Google DeepMind released WeatherNext 3 on 3 September — a forecasting model at 5 km resolution for temperature and humidity, refreshed hourly instead of every six hours. The biggest jump is in precipitation: up to 60% CRPS improvement over WeatherNext 2. It is going into Google Search, Maps and the Gemini app.
Key takeaways
- Resolution of 5 km for temperature and humidity, 10 km for other surface variables, 25 km for atmospheric wind
- Hourly refresh instead of the six-hour increments in WeatherNext 2
- Precipitation: up to 60% CRPS improvement verified against IMERG, 30% MRMS, 10% rain gauges
- It ingests real-time satellite mosaics, NASA's IMERG data and sparse ground-station observations
- Access via Search, Gemini, Google Maps, the Maps Platform Weather API, Earth Engine, BigQuery and Cloud Storage
Three resolutions instead of one
Google dropped the uniform grid. Across the three scales combined, that makes the picture roughly five times sharper. The cadence changed too: forecasts refresh hourly rather than every six hours, which matters for fast-moving phenomena.
| Variable | Grid resolution |
|---|---|
| Temperature and humidity | 5 km |
| Other near-surface variables | 10 km |
| Wind in the upper atmosphere | 25 km |
Precipitation, the hardest variable
Rain has traditionally been the weak spot of weather models. Google gives three numbers, each against a different reference: up to 60% CRPS improvement when verified with satellite IMERG data, 30% with radar-based MRMS and 10% with rain-gauge measurements. The spread is not incidental — the denser and more direct the measurement network, the smaller the model's edge.
It also claims up to 50% better precipitation forecasts a day ahead, especially in regions with thin measurement infrastructure.
Where the data comes from
WeatherNext 3 combines four streams: hourly mosaics from geostationary satellites, NASA's IMERG precipitation data, a global precipitation reanalysis based on satellite radar, and sparse ground-station observations that pin down local topography. Winds at 100 metres — turbine height — plus cloud cover and solar radiation are computed separately, aimed at renewable energy.
Why it matters
Weather forecasting is one of the few AI applications where quality is checked daily and in public. Moving to hourly refresh and lifting precipitation accuracy feeds straight into planning in agriculture, logistics and wind power.
What's next
- Independent measurement is run by Brightband, which publishes results on its own live leaderboards
- It enters Google Search, the Gemini app and Google Maps, with developer access through the Maps Platform Weather API, Earth Engine, BigQuery and Cloud Storage
