WeatherNext 3 DeepMind: The Research Behind the Model
WeatherNext 3 is a Google DeepMind research project that reached ordinary users through Google's own products. This page covers the organization behind it, where it sits in the research line, and which sources to read if you want the model's own claims instead of second-hand summaries. For the technical picture, see the model overview; for the write-up itself, see the key points of arXiv 2609.03582.
7-day free trial, then $19/mo · cancel anytime · compare plans
Three anchors of the project
Where the model comes from, in the shortest possible form.
Google DeepMind The research organization behind the model, and the home of the earlier weather systems that led to it.
The research organization behind the model, and the home of the earlier weather systems that led to it.
3rd generation WeatherNext 3 is the third generation of a named line, following WeatherNext and WeatherNext 2, each one documented publicly.
WeatherNext 3 is the third generation of a named line, following WeatherNext and WeatherNext 2, each one documented publicly.
Product reach Output feeds Google Search, Gemini and Maps, which is why most people meet the model without ever signing up for anything.
Output feeds Google Search, Gemini and Maps, which is why most people meet the model without ever signing up for anything.
How the research line got here
Each step in the family answered a question the previous step left open. Reading the line in order makes the claims in the latest generation easier to judge.
GraphCast
An earlier learned global model from the same team, widely cited as the moment machine learning forecasting stopped being a curiosity and started being compared with operational systems.
GenCast
The probabilistic successor: instead of a single answer, an ensemble of plausible futures. It set the pattern the later generations still follow.
WeatherNext
The naming arrived with the product era - a generation built to be read by people and pipelines, not only evaluated by researchers.
WeatherNext 3
Hourly global output, sharper local detail, raw satellite observations taken in directly, and daily use inside Google's own products.
What the team actually works on
Behind a launch note there is a working list. These four threads explain most of what changed in this generation.
Data ingestion
Turning raw geostationary satellite observations into model-ready inputs - the change that defines this generation more than any other.
Forecasting skill
Training and scoring ensembles with CRPS against named baselines such as ECMWF ENS and AIFS ENS, rather than against nothing.
Precipitation
Satellite-derived precipitation estimation at 0.1 degree, the area where Google's own evaluation reports the largest relative gain.
Delivery
Working with product teams to place forecasts in Search, Gemini and Maps, and with platform teams that meter access to the data products.
How to check the research yourself
Research pages are the primary source; summaries, including this one, are not. These are the routes worth using.
Read the write-up
The technical report documents the method, the training data and the evaluation protocol behind the headline numbers. Start there before quoting anything.
Look at the code
Parts of the family are published in public repositories - model code, weights for earlier generations and notebooks - so claims can be inspected, not just read.
Use the research surface
Weather Lab exposes forecast layers interactively, which is the closest thing to standing inside the research without running anything locally.
Separate research from product
A published metric and a product rollout are different claims with different evidence. Keeping them apart is most of a good evaluation.
WeatherNext 3 DeepMind questions
Short answers about the organization, the family and what is actually published.
Use the results without running the research
Hourly forecasts, local detail and one workspace, starting with a 7-day free trial. Background: the paper notes and what is published on GitHub.
Related guides
The pages that go with this one, chosen by what each one answers.
