Google’s WeatherNext 3 model now incorporates satellite observations directly, reducing its dependence on global atmospheric snapshots known as reanalyses. Because those snapshots are generally produced every six hours, adding lower-latency observations lets the system generate forecasts hourly.
Google also increased the model’s spatial resolution and added a separate system trained on satellite precipitation estimates. The company reports about a 5 percent improvement in upper-atmosphere accuracy over WeatherNext 2, equivalent in its analysis to roughly six additional hours of useful lead time. A location-specific method using land or ocean status and elevation improved some surface-temperature predictions by as much as 30 percent.
The white paper also documents limitations. WeatherNext 3 performs worse than comparison models on several variables at the initial six-hour horizon before pulling ahead later, and some precipitation maps reveal hexagonal grid patterns. Its generated temperature scenarios can also shift the global average when ideally local variation would cancel out. The model now supplies weather information across Google Search, Gemini and Maps, making those remaining quirks relevant to consumer forecasts.