NVIDIA open-sources Earth-2 weather models for nowcasting, 15-day forecasts, and data assimilation

NVIDIA is expanding its Earth-2 lineup with three open models that cover the core steps of modern weather prediction—starting conditions, short-term storm evolution, and medium-range global forecasts—packaged so teams can run AI weather pipelines on their own infrastructure.

Open models that cover nowcasting and medium-range forecasting

For short time horizons, NVIDIA has published Earth-2 Nowcasting on Hugging Face, built on a new architecture called StormScope. The model generates zero- to six-hour, kilometer-scale predictions designed to capture local storms and hazardous weather quickly, by directly predicting satellite and radar-like data instead of relying purely on physics simulations. The released checkpoint is trained on geostationary satellite observations (GOES) over the contiguous United States, with NVIDIA noting the approach can be adapted to other regions with comparable satellite coverage. Details are in the accompanying paper, “Learning Accurate Storm-Scale Evolution from Observations” (https://research.nvidia.com/publication/2026-01_learning-accurate-storm-scale-evolution-observations) and the model card (https://huggingface.co/nvidia/stormscope-goes-mrms).

For planning windows that matter to logistics, demand forecasting, and energy-aware operations, Earth-2 Medium Range targets up to 15-day global forecasts across more than 70 variables (temperature, pressure, wind, humidity). It uses an Atlas latent diffusion transformer that predicts incremental atmospheric changes—an approach meant to preserve large-scale structures and reduce drift over time. NVIDIA claims benchmark wins versus leading open models on common variables. Reference links: paper (https://research.nvidia.com/publication/2026-01_demystifying-data-driven-probabilistic-medium-range-weather-forecasting) and checkpoint (https://huggingface.co/nvidia/atlas-era5).

Data assimilation plus open tooling for sovereign pipelines

A third component, Earth-2 Global Data Assimilation (HealDA), is described as an end-to-end model that produces “initial conditions” (the global snapshot of temperature, winds, humidity, and pressure used to start forecasts). NVIDIA says it can generate these in seconds on GPUs instead of hours on supercomputers, and that pairing HealDA with Medium Range yields an entirely AI-based open forecasting pipeline. HealDA is “coming soon” to Hugging Face; research link: https://research.nvidia.com/publication/2026-01_healda-highlighting-importance-initial-errors-end-end-ai-weather-forecasts.

To run inference and assemble workflows, NVIDIA points developers to Earth2Studio (https://github.com/NVIDIA/earth2studio). For training and fine-tuning, it highlights Physics Nemo (https://github.com/NVIDIA/physicsnemo). For B2B teams, the operational takeaway is control: open checkpoints and open tooling enable “sovereign” weather intelligence that can be customized with proprietary data and deployed where compliance requires.