NVIDIA launches Earth-2 open models for AI weather forecasting
Company introduces what it calls a fully open, GPU-accelerated stack for AI-driven weather and climate forecasting.
NVIDIA has unveiled a new Earth-2 family of open models and tools for weather and climate AI, positioning it as a fully open, accelerated software stack that developers and forecasters can run on their own infrastructure.
According to the company, the release is aimed at speeding up forecasting, lowering compute costs and making state-of-the-art capabilities accessible to agencies, enterprises and startups that do not operate large supercomputers.
What the Earth-2 stack includes
Earth-2 comprises pretrained models for global medium-range forecasting and high-resolution nowcasting, along with components for data assimilation and AI-based downscaling. It is complemented by open libraries for data handling, inference and evaluation, plus recipes for fine-tuning on local datasets. NVIDIA says the stack is designed for end-to-end workflows, from ingesting observations to producing city-scale guidance, and that models can be combined or swapped depending on a user’s operational needs.
As per the company, the models run efficiently on GPUs, delivering significant speedups versus conventional numerical weather prediction. That efficiency could enable more frequent updates, larger ensembles and higher spatial resolution, areas that are often constrained by time and cost on traditional systems. The emphasis on openness, including access to model weights and reference implementations, is intended to help researchers reproduce results and adapt methods for regional conditions.
Why this matters for India and emerging markets
Extreme weather is intensifying across South Asia, with urban flooding, heatwaves and cyclones affecting lives and livelihoods. Faster and more localised guidance can help utilities manage renewable generation, insurers assess exposure and city authorities plan responses.
The company highlights use cases across energy, agriculture, logistics and disaster management. For example, nowcasting at kilometre-level resolution can inform short-term grid balancing for wind and solar, while AI downscaling can translate global signals into neighbourhood-scale rain and wind estimates. Financial services firms can also use rapid ensembles to stress-test portfolios against cyclone tracks or heat extremes.
How can developers access and customise the models
Developers can obtain the open models through the Earth-2 tooling and community repositories, then deploy on workstations, on-prem clusters or cloud instances. According to the company, teams can fine-tune with local observations, integrate radar or satellite feeds, and build pipelines that chain data assimilation, global forecasts and downscaling into one service. Standardised APIs and evaluation suites are included so that agencies and startups can compare methods on common datasets and track improvements over time.
NVIDIA founder and CEO Jensen Huang has said the company is investing in open model families to accelerate scientific computing and climate resilience. With Earth-2, NVIDIA is signalling a commitment to open access, reproducibility and developer tooling, while leveraging its GPU platform for performance. The company says further releases will expand the library of models and datasets, add more reference pipelines and deepen support for evaluation in operational settings.