Google's WeatherNext 2 ships as forecast data across Earth Engine, BigQuery, and Vertex AI

What the model claims, precisely

Google DeepMind and Google Research say WeatherNext 2 generates forecasts 8x faster than its predecessor and at resolution down to one hour. From a single starting point it produces hundreds of possible weather scenarios, and the announcement states each prediction takes less than a minute on a single TPU.

The comparison claim is unusually specific: WeatherNext 2 surpasses the previous WeatherNext model on 99.9% of variables — temperature, wind, humidity — and lead times spanning 0 to 15 days. That framing is a like-for-like beat against Google's own prior state of the art, not a claim against physics-based systems, though the post notes the equivalent physics-based ensemble would take hours on a supercomputer.

The marginals-to-joints trick

The technical core is what Google calls a Functional Generative Network (FGN), which injects noise directly into the model architecture so generated forecasts stay physically realistic and interconnected. The interesting design choice is what the model is trained on versus what it produces.

According to the post, the model is trained only on 'marginals' — standalone quantities like temperature at a point or wind speed at an altitude — yet learns to forecast 'joints,' the interconnected systems where all those pieces fit together. Google frames this as the capability behind its most useful outputs, such as identifying entire regions under high heat or expected power output across a wind farm.

What's novel about our approach is that the model is only trained on these marginals. Yet, from that training, it learns to skillfully forecast 'joints' — large, complex, interconnected systems that depend on how all those individual pieces fit together.Montana Labs

Distribution is the actual news

The announcement doesn't stop at model quality. WeatherNext 2 forecast data is now in Earth Engine and BigQuery, with an early access program on Vertex AI for custom model inference. That means developers can query forecasts through a data warehouse and a geospatial catalog rather than standing up their own inference stack.

Simultaneously, Google says the technology now upgrades weather in Search, Gemini, Pixel Weather, and the Maps Platform Weather API, with Google Maps to follow in the coming weeks. So the same research is being routed to two audiences at once: builders who consume it as data, and consumers who never see the model name.

The specific implication: cheap ensembles change who can afford probabilistic forecasting

The line worth underlining is 'hundreds of possible weather outcomes' for 'less than a minute on a single TPU.' Probabilistic ensemble forecasting has historically been gated by supercomputer time, which is precisely why the worst-case scenarios that matter most for planning are expensive to generate at scale.

By collapsing that cost and delivering the output through BigQuery and Earth Engine, Google is putting scenario-based forecasting within reach of teams that could never run a physics-based ensemble — supply chain planners, energy operators, logistics companies. The open question the post leaves unanswered is pricing, latency guarantees, and update cadence for these endpoints, which is what will determine whether this is a usable operational feed or a research preview with a query interface.