Google DeepMind has launched WeatherNext 2, a new AI weather model that Google says produces hundreds of possible forecast scenarios from a single input in under one minute on a single TPU. The company also says it runs 8 times faster than the original WeatherNext and beats it on 99.9% of variables and lead times. Those accuracy numbers are Google’s comparison with its own previous model, not an independent test against national weather services.
What happened
Google announced WeatherNext 2 on 17 November in a post on its official blog. The headline claim is that the new model “surpasses our previous state-of-the-art WeatherNext model on 99.9% of variables and lead times”. Put simply, across almost every quantity it forecasts, and across almost every forecast horizon, Google measured the new version doing better than the old one.
Speed is the second part of the story. Google says WeatherNext 2 generates forecasts 8 times faster than its predecessor, with resolution of up to 1 hour. Google’s Yossi Matias highlighted the same figures in a post on X on launch day.
Unlike many research announcements, this one is already in products. Google says WeatherNext 2 has rolled out in Search, Gemini and Pixel Weather, and is coming to Google Maps. For developers and scientists, the forecast data is available in Earth Engine and BigQuery, with early access on Vertex AI.
How it works
Weather forecasting has a basic problem: the atmosphere is chaotic. A tiny difference in today’s conditions can grow into a very different storm track five days from now. So serious forecasters do not rely on one prediction. They run many slightly different versions and look at the spread. If most versions agree on heavy rain, confidence is high. If they scatter, the honest answer is uncertainty.
Traditionally, running those many versions has meant a lot of time on large supercomputers. The interesting part of WeatherNext 2 is how cheaply Google says it can produce them: hundreds of scenarios in under a minute on one TPU, the kind of AI chip Google builds.
The technique behind it is what Google calls a Functional Generative Network, or FGN. In Google’s words, it “injects ‘noise’ directly into the model architecture so the forecasts it generate remain physically realistic and interconnected”. Here is how I think about it. Picture a skilled jazz musician asked to play the same tune a hundred times. Each version is different, but every one stays in key and follows the song. The noise is what makes each run different. Building it into the model itself, rather than bolting it on afterwards, is meant to keep each scenario coherent, so temperature, wind and pressure in one forecast still make sense together.
By the numbers
| Item | Figure | Source |
|---|---|---|
| Variables and lead times where it beats the previous WeatherNext | 99.9% | Google blog, 17 Nov 2025 |
| Speed compared with predecessor | 8 times faster | Google blog and Yossi Matias, 17 Nov 2025 |
| Time resolution | Up to 1 hour | Google blog and Yossi Matias, 17 Nov 2025 |
| Scenarios from a single input | Hundreds, in under 1 minute | Google blog, 17 Nov 2025 |
| Hardware for that run | 1 TPU | Google blog, 17 Nov 2025 |
| Products using it now | Search, Gemini, Pixel Weather | Google blog, 17 Nov 2025 |
Why it matters
What caught my attention is not the 99.9% figure. Beating your own previous model is useful, but it is also the easiest comparison to win. The part I find more significant is the cost of uncertainty. If you can generate hundreds of scenarios in under a minute on a single chip, you can afford to ask “how sure are we?” far more often and at finer detail. That is exactly the question that matters when a storm or a heatwave is approaching.
The distribution is the other big shift. Because it is going into Search, Gemini and Pixel Weather, faster and more detailed forecasts could reach a very large number of phone users who will never know which model produced them. For those of us in hot parts of the world, better short range warnings about heat are not an abstract improvement.
I would still keep two cautions in mind. First, all the accuracy claims here come from Google itself, measured against its own earlier model, and there is no independent comparison in the announcement with every national weather service. Second, a phone app forecast is only as useful as the way it communicates uncertainty. Having hundreds of scenarios helps only if people see what that spread means.
What comes next
Google says WeatherNext 2 is coming to Google Maps. On the research side, the data is already available in Earth Engine and BigQuery, and early access is open on Vertex AI, so outside scientists and companies can start testing it for themselves. Those independent tests are what I will be watching most.
Sources
- Google blog, Google DeepMind announcement of WeatherNext 2, 17 November 2025, https://blog.google/technology/google-deepmind/weathernext-2/
- Yossi Matias on X, post announcing WeatherNext 2, 17 November 2025, https://x.com/ymatias/status/1990436360717516854