---
id: engines/weathernext-nvidia-earth-2
about:
  label: WeatherNext / NVIDIA Earth-2
  description: >-
    Two AI weather-forecasting stacks: Google DeepMind's WeatherNext 3, a hosted
    global model that forecasts hourly and feeds Google products and cloud
    datasets, and NVIDIA's Earth-2, a family of open models and tools that
    organisations run on their own GPUs.
  tags:
    - weather forecast
    - AI model
    - nowcasting
    - downscaling
    - open models
tool:
  form: Forecast models and services
  group: Weather and climate
  homepage: 'https://deepmind.google/science/weathernext/'
open:
  source: partly
  data: open
  api: 'yes'
  participation: unknown
domains:
  - climate
relations:
  - predicate: answers
    target: questions/weather-and-climate
provenance:
  origin: agent
  sources:
    - title: WeatherNext 3 — Google DeepMind
      url: 'https://deepmind.google/science/weathernext/'
    - title: 'WeatherNext 3: Our most advanced global weather AI model'
      url: >-
        https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-weathernext-3/
    - title: >-
        [2609.03582] WeatherNext 3: Increasing resolution and performance of
        global weather models with raw observations
      url: 'https://arxiv.org/abs/2609.03582'
    - title: NVIDIA Earth-2
      url: 'https://www.nvidia.com/en-us/high-performance-computing/earth-2/'
    - title: >-
        NVIDIA Launches Earth-2 Family of Open Models — the World's First Fully
        Open, Accelerated Set of Models and Tools for AI Weather | NVIDIA Blog
      url: 'https://blogs.nvidia.com/blog/nvidia-earth-2-open-models/'
  reviewed: '2026-10-18'
  reviewed_by: 'Claude, from the sources listed'
  status: draft
import:
  source: openai-linkedin-scrape 2026-10-17
  id: DT017
  kind: forecast_model
  family: models_simulations_and_forecasts
  priority: high
  verified: true
  posts:
    - P003
    - P131
    - P146
  posters:
    - Yossi Matias
    - Emma Scharfmann
    - David Carlin
---
WeatherNext 3 is Google DeepMind's global weather model, announced on 3 September 2026 with a paper by Rasp and 24 co-authors. It is an ensemble model that ingests low-latency geostationary satellite data and issues a new forecast every hour rather than every six, at 0.1 degree resolution for surface variables including solar radiation and cloud cover, and it learns directly from weather-station observations so it can predict 2 m temperature and dewpoint at any location on a 5 km grid, including places with no station. It also predicts satellite-derived precipitation and tropical cyclones. Google's page cites its use by the US National Hurricane Center for Hurricane Melissa's landfall in Jamaica.

It is a hosted Google service, not a model to download. Forecasts feed Google Search, Maps, the Gemini app, the Maps Platform Weather API and Earth Engine, and the forecast data can be queried in BigQuery and Earth Engine or bulk-downloaded from Google Cloud Storage without any model setup. Weather Lab is an experimental site for viewing the forecasts and tracking tropical cyclones.

NVIDIA Earth-2 is a family of open models, libraries and frameworks for weather and climate AI, announced on 26 January 2026 at the American Meteorological Society meeting. The models cover the forecasting chain: Earth-2 Global Data Assimilation (HealDA) builds initial conditions from observations in seconds on GPUs; Earth-2 Medium Range (Atlas architecture) forecasts over 70 variables up to 15 days ahead; Earth-2 Nowcasting (StormScope) makes kilometre-scale zero-to-six-hour storm forecasts by predicting satellite and radar imagery; CorrDiff downscales coarse forecasts to regional resolution; and FourCastNet3 is a global ensemble model. Earth-2 also wraps open models from ECMWF, Microsoft and Google. NVIDIA's announcement names Brightband running Medium Range daily, the Israel Meteorological Service running CorrDiff operationally, and TotalEnergies, Eni, GCL, Southwest Power Pool with Hitachi, The Weather Company and the US National Weather Service evaluating or using the models.

Earth-2 is built to run on an organisation's own GPU infrastructure: the models are on GitHub and Hugging Face, Earth2Studio is the toolkit for building, fine-tuning and deploying them, PhysicsNeMo is the open-source Python framework for training, and the maker also offers CorrDiff as a NIM microservice and an Omniverse blueprint for visualisation.

Status: both are operational as of the dates above.