Google DeepMind Introduces WeatherNext 3, an Hourly Global AI Weather Model
Google DeepMind
Google DeepMind and Google Research announced WeatherNext 3, a global AI weather model that learns from live satellite observations and issues hourly forecasts at up to 5-kilometer resolution, roughly five times sharper than WeatherNext 2. Google says it is rolling the model into Search, Gemini, Maps, Google Maps Platform and Earth Engine, and publishing forecast data via BigQuery and Cloud Storage.
What Google announced
Google DeepMind and Google Research introduced WeatherNext 3, described as their most advanced global weather model to date. Per the announcement, the key architectural change is the training data: instead of learning mainly from numerical weather prediction outputs that carry a six-hour lag, the model ingests a mosaic of live geostationary satellite data alongside historical analysis. That supports a forecast refreshed every hour at multiple resolutions - about 5 kilometers for some surface variables, 10 kilometers for others and 25 kilometers for atmospheric variables - which Google says makes the global picture roughly five times sharper than WeatherNext 2. The model also predicts turbine-height (100-meter) wind speeds, cloud cover and solar radiation, and is trained on NASA IMERG and Google's satellite-radar precipitation reanalysis. Google reports CRPS improvements of up to 60% against IMERG, 30% for MRMS and 10% against rain gauges at early lead times, and says consumer precipitation forecasts a day or more ahead are up to 50% more accurate. Google states the model is being integrated into Search, Gemini, Maps, Google Maps Platform and Earth Engine, with data queryable in BigQuery and downloadable from Cloud Storage.
A scenario for a grid operations analyst
A renewable energy analyst at a grid operator could consider testing WeatherNext 3's 100-meter wind and solar radiation fields as an additional input alongside existing numerical weather prediction feeds. Because the model refreshes hourly and resolves finer spatial detail, the plausible use case is short-horizon planning - for example, comparing day-ahead wind generation estimates against the operator's current source. This remains a conditional experiment: the announcement describes the variables and availability channels but does not establish that the model outperforms any specific incumbent system in a given region or that integration into a production dispatch process is warranted.
How to run a bounded trial
Before adopting anything, a team could define a baseline: pick a region and a fixed historical window, and record the incumbent forecast source's errors for wind speed and precipitation. Then pull the same periods from WeatherNext 3 via the described Cloud Storage or BigQuery access and compute comparable error metrics using identical verification stations. Success criteria should be set in advance, such as matching or beating baseline error without larger bias in extreme cases. Failure checks should include fast-changing events like convective storms and coastal or mountainous terrain, where the announcement claims the largest gains. A meteorologist should review discrepancies, and official warnings from the local meteorological agency should remain the authoritative source for safety decisions regardless of trial outcomes.
Trade-offs and open questions
The finer grid and hourly cadence imply larger data volumes, so storage, query cost and pipeline engineering are prerequisites worth scoping first. The reported benchmark gains are Google's own evaluations against named baselines and may not transfer to every variable, season or region. The excerpt does not detail uncertainty quantification, latency, licensing terms or how the model handles rare extreme events, and it explicitly warns that the atmosphere retains unpredictability. The page also directs users to national meteorological services for official forecasts and warnings. A team should not rely on this model for severe-weather alerting, regulatory reporting or safety-critical decisions, and should treat it as a complementary input pending independent validation.
Context and limitations
- Based on a short source excerpt; the full article and research paper were not analyzed.
- Accuracy figures are Google's own reported evaluations against named baselines, not independent replication.
- The page directs users to local meteorological agencies for official warnings and safety advisories.
- Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.