Google released WeatherNext 3, an updated machine-learning weather model that directly ingests some satellite observations instead of relying entirely on six-hour global reanalysis snapshots. The added data allow the system to produce forecasts hourly, reducing the delay between new observations and a generated outlook.
The model also runs at higher spatial resolution and includes a separate machine-learning system trained on satellite-based precipitation estimates. For surface temperature and dew point at a requested location, Google added limited physical descriptors such as land or ocean, elevation and historical station information. Most of the forecast remains learned from past atmospheric patterns rather than calculated through a full physics simulation.
According to the company’s white paper, upper-atmosphere accuracy improved by roughly 5 percent over WeatherNext 2, which the team equates to about six additional hours of useful forecast lead time. The point-location surface-temperature method improved some results by as much as 30 percent. Ars Technica noted that those figures are reported by the model’s developers.
The evaluation also identifies weaknesses. For several variables, WeatherNext 3 performed worse in the initial six-hour forecast before improving later in the 15-day window. Some precipitation maps display hexagonal grid patterns, and ensemble surface-temperature outputs can produce global averages that shift higher or lower instead of balancing local variation.
Google is using WeatherNext 3 for forecast information in Search, Gemini and Maps. The public paper does not settle how the model performs during rare extremes, in regions with sparse observations or against every national forecasting system. Continued verification against real events and independent benchmarks will show where the update is dependable.
The checked record for Google’s WeatherNext 3 Adds Satellite Data and Hourly AI Forecasts establishes the following dated facts: WeatherNext 3 ingests some raw satellite observations. The model can generate forecasts hourly. Google added higher spatial resolution and a separate precipitation model. The paper reports about a 5 percent upper-atmosphere accuracy gain. Some initial six-hour forecasts performed worse than comparison models. The relevant background is also specific: Reanalysis products combine many observations into six-hour global snapshots. Machine-learning weather models use much less compute than many physics models. Google now uses the model in Search, Gemini and Maps. Performance claims come from the developer paper, and short-range weaknesses and grid artifacts remain. The cited reporting does not resolve questions beyond those stated limits.