IBM and NASA open-source a climate foundation model

Prithvi WxC is an open-source weather and climate foundation model developed by IBM and NASA, with contributions from Oak Ridge National Laboratory. You can download the model weights and code right now on Hugging Face.

Instead of acting as a narrow, single-purpose forecasting tool that only tells you tomorrow's temperature, Prithvi WxC serves as a versatile backbone for Earth system science. It can handle global-scale simulations just as easily as local-scale downscaling.

What it does

The model is pre-trained on 40 years of Earth observation data from NASA's MERRA-2 reanalysis dataset. Its architecture is explicitly designed so scientists can fine-tune it across global, regional, and local contexts.

Alongside the base model, IBM and NASA released two fine-tuned models on Hugging Face targeting specific research challenges.

The first fine-tuned model focuses on weather and climate data downscaling. It can enhance the spatial resolution of inputs like temperature, precipitation, and surface winds by up to 12x, enabling localized forecasting and granular long-term climate projections.

The second fine-tuned model tackles gravity wave parameterization. Atmospheric gravity waves drive key processes like cloud formation and aircraft turbulence, yet traditional numerical climate models struggle to capture them. Prithvi WxC helps scientists estimate gravity wave generation with much higher fidelity.

Real-world testing is already underway. Environment and Climate Change Canada is experimenting with the model for precipitation nowcasting from real-time radar data, as well as downscaling 15-kilometer global forecasts down to kilometer-scale resolution.

In paper benchmarks published on arXiv, the foundation model was able to accurately reconstruct global surface temperatures from a random sample of just 5 percent of original data, proving its strength in data assimilation.

How to get started

You can access the base foundation model and the gravity wave parameterization model directly through the NASA-IBM page on Hugging Face. The downscaling model is hosted on the IBM Granite page on Hugging Face.

Researchers looking for architectural details and experimental setups can check out their paper, titled Prithvi WxC: Foundation Model for Weather and Climate, on arXiv.

Why it matters

If you are a climate scientist, an environmental data developer, or an enterprise dealing with localized climate risk, this is worth exploring right away. Pre-training on 40 years of NASA data requires immense compute, and having that foundation open-sourced removes massive technical barriers.

We think IBM and NASA took the smarter path here. Rather than building another proprietary, black-box forecast product, they gave the scientific community an open, multi-purpose platform that addresses real physical modeling bottlenecks like gravity waves and resolution limits.