WeatherNext 2
Last reviewed
Jun 7, 2026
Sources
13 citations
Review status
Source-backed
Revision
v1 · 1,793 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 7, 2026
Sources
13 citations
Review status
Source-backed
Revision
v1 · 1,793 words
Add missing citations, update stale details, or suggest a clearer explanation.
WeatherNext 2 is an artificial intelligence weather forecasting model from Google DeepMind and Google Research, announced on November 17, 2025. It is the second generation of the company's WeatherNext family and succeeds the original WeatherNext, which packaged the earlier GraphCast and GenCast research models for production use. WeatherNext 2 is built on a new modeling approach that Google DeepMind calls a Functional Generative Network (FGN). The model produces large ensembles of probabilistic forecasts out to 15 days, generating each scenario in under a minute on a single Tensor Processing Unit (TPU), roughly eight times faster than the prior model. Google reports that WeatherNext 2 outperforms its predecessor on 99.9% of measured variables and lead times. The model is distributed through Google Cloud (Vertex AI, BigQuery) and Earth Engine, and it powers upgraded weather forecasts in Google Search, Google Maps, Pixel Weather, and Gemini.
WeatherNext 2 is the latest step in a lineage of machine learning weather models developed at Google DeepMind as part of its AI for science work. The first major model, GraphCast, was published in 2023. GraphCast is a deterministic graph neural network that produces a single medium-range forecast out to 10 days at 0.25 degree resolution, and it demonstrated that a trained neural network could match or beat the deterministic output of leading numerical weather prediction (NWP) systems while running far faster.
GraphCast produces only one forecast per run, which limits its usefulness for assessing uncertainty and rare, high-impact events. To address this, Google DeepMind introduced GenCast, published in the journal Nature on December 4, 2024. GenCast is a diffusion model adapted to the sphere of the Earth. Rather than emitting a single answer, it generates an ensemble of 50 or more possible weather trajectories at 0.25 degree resolution out to 15 days, with each member representing one plausible future. In a head to head evaluation against ENS, the operational ensemble run by the European Centre for Medium-Range Weather Forecasts (ECMWF), GenCast was more accurate on 97.2% of 1,320 variable and lead-time combinations, rising to 99.8% at lead times beyond 36 hours. A single Cloud TPU v5 produced a 15-day GenCast forecast in about eight minutes.
In 2025 Google folded these research models into a product family called WeatherNext. WeatherNext Graph is the operational version of the GraphCast model, and WeatherNext Gen is the operational version of GenCast. Google Cloud announced enterprise availability of the WeatherNext models on March 5, 2025, offering them through the Vertex AI Model Garden and through experimental datasets in BigQuery and Earth Engine, aimed at customers in energy, retail, insurance, and logistics. WeatherNext 2 is the successor to this WeatherNext Gen model.
The technical foundation of WeatherNext 2 is the Functional Generative Network, described in the paper "Skillful joint probabilistic weather forecasting from marginals," posted to arXiv (2506.10772) on June 12, 2025 by Ferran Alet, Ilan Price, and colleagues at Google DeepMind.
A Functional Generative Network takes a single starting state of the atmosphere and a small random input, a 32-dimensional Gaussian noise vector drawn fresh at each step, and maps them to a complete future weather field. The noise is injected directly into the network's parameters as a learned model perturbation rather than being added to the data, so each random draw effectively yields a slightly different but self-consistent model. Running the network many times with different noise vectors produces an ensemble of forecasts. The published configuration uses a transformer with about 180 million parameters per model seed, a latent dimension of 768, and 24 layers, run on a 0.25 degree latitude-longitude grid at a 6-hour timestep, covering 6 atmospheric variables at 13 pressure levels plus 6 surface variables. The released system combines four independently trained seeds.
The key idea, reflected in the paper's title, is that FGN is trained only on "marginals," the forecast distribution at each individual location and variable, by directly minimizing the Continuous Ranked Probability Score (CRPS). Despite never being trained on the "joints," the way variables across space combine into coherent structures such as cyclones, fronts, and heat waves, the model still captures that joint spatial structure when sampled. Google DeepMind frames this as letting WeatherNext 2 sample directly from the joint distribution over 15-day global weather trajectories while keeping training simple and stable.
This distinguishes FGN from the diffusion-based GenCast. GenCast generates an ensemble member through an iterative denoising process, gradually turning random noise into a weather field over many refinement steps, which is accurate but computationally heavy. FGN instead perturbs the model itself and produces a forecast field in a single forward pass per member, which is the main source of its speed advantage. Both are generative and probabilistic, but they reach an ensemble by different mechanisms: diffusion over the output for GenCast, learned perturbation of the network for FGN.
Google reports that WeatherNext 2 surpasses the previous WeatherNext (GenCast-based) model on 99.9% of combinations of variable, pressure level, and lead time, across all lead times from 0 to 15 days. The variables include temperature, wind, and humidity. In the FGN paper, this corresponds to better CRPS in 99.9% of cases with an average CRPS improvement of about 6.5%.
On tropical cyclones, the FGN paper reports that the model's cyclone track position errors correspond to roughly one extra day of useful predictive skill compared with GenCast, meaning its forecast at a given lead time is about as accurate as GenCast's forecast a day earlier. Because GenCast already outperformed the ECMWF ENS ensemble on the large majority of targets, Google positions WeatherNext 2 as a further improvement on top of that baseline rather than restating a direct ENS comparison for the new model. Earlier WeatherNext cyclone evaluations against NHC observations from 2023 and 2024 found 5-day track predictions averaging about 140 km closer to the true cyclone location than ECMWF ENS, an improvement of roughly 1.5 days of lead time.
These figures come from Google and Google DeepMind and from the company's own arXiv paper, and they are measured largely against the ERA5 reanalysis archive that the models were trained on, so independent verification by external groups and operational agencies remains important. As context, during the 2025 Atlantic hurricane season the U.S. National Hurricane Center reported that a WeatherNext cyclone model was among the top-performing individual models for track and intensity, and Google DeepMind described the model anticipating Hurricane Melissa's rapid intensification and landfall in Jamaica about five days in advance.
| Model | Type | Approach | Ensemble | Lead time | Resolution | Speed (15-day) | First public |
|---|---|---|---|---|---|---|---|
| GraphCast | Deterministic | Graph neural network | 1 forecast | up to 10 days | 0.25 deg | minutes on one TPU | 2023 |
| GenCast / WeatherNext Gen | Probabilistic | Diffusion on the sphere | 50+ members | up to 15 days | 0.25 deg | about 8 min on TPU v5 | Dec 4, 2024 (Nature) |
| WeatherNext 2 | Probabilistic | Functional Generative Network | hundreds of members | up to 15 days | 0.25 deg, hourly outputs | under 1 min per member on one TPU, about 8x faster | Nov 17, 2025 |
WeatherNext 2 is available to developers and enterprises through several Google platforms. Forecast datasets are published in Earth Engine and BigQuery for analysis, and an early access program on Google Cloud's Vertex AI allows organizations to run custom inference with the model. Google describes the efficiency gain, hundreds of scenarios generated in under a minute on a single TPU, as what makes it practical to explore far more possible outcomes for risk-sensitive users such as energy traders, insurers, and emergency planners. The model can also produce hour-by-hour outputs, a finer temporal resolution than the 6-hour and 12-hour steps of the earlier models.
On the consumer side, Google integrated WeatherNext 2 into the core forecasting system behind its weather features. It powers upgraded forecasts in Google Search, the Gemini assistant, Pixel Weather on Pixel phones, and the Google Maps Platform Weather API, with broader use across Google Maps rolling out after launch. An experimental cyclone-focused version of WeatherNext 2 runs in Weather Lab, Google DeepMind's site for sharing models with meteorological agencies and emergency services, where it predicts cyclone path, intensity, structure, size, and formation locations up to 15 days ahead.
WeatherNext 2 reflects a broader shift in AI weather forecasting from deterministic single forecasts toward fast, large probabilistic ensembles produced by neural networks rather than physics-based supercomputer simulations. By generating each ensemble member in a single forward pass, the FGN approach makes very large ensembles cheap enough to run routinely, which matters most for low-probability, high-impact events where the spread of outcomes is the decision-relevant quantity. The technical contribution, training only on per-location marginals yet recovering coherent joint spatial structure, is also of interest beyond weather as a general pattern for probabilistic deep learning.
At the same time, these systems do not replace operational forecasting outright. They are trained on and benchmarked against reanalysis data, they still depend on the observation and data-assimilation pipelines run by agencies such as ECMWF and NOAA to define their starting conditions, and most of the headline performance numbers originate from the developer rather than from independent evaluators. The collaboration with the National Hurricane Center during the 2025 season illustrates the intended role: AI models such as WeatherNext 2 as additional, fast, skillful inputs that human forecasters and physics-based models weigh alongside one another.