NVIDIA Earth-2 launches open models for weather | Keryc
NVIDIA has just put on the table an open collection of models designed to cover the entire weather stack. Can you imagine generating initial conditions, making 15-day forecasts, and producing kilometer-scale nowcasts in minutes, all with open tools accelerated by GPUs? That’s exactly what Earth-2 proposes.
What Earth-2 is and why it matters
Earth-2 is a set of open source models and tools that aims to unite capabilities that until now were fragmented: data assimilation, forecasting, nowcasting, downscaling and more. The idea is to let developers, institutions and countries build their own prediction chains using their data and infrastructure. Why is that useful? Because operational meteorology is not just science: it’s sovereignty over data and decisions.
Earth-2 enables you to create sovereign, customizable meteorological capabilities by training and running models on your own GPUs.
Key models and their architectures
Earth-2 Nowcasting (StormScope): designed for 0–6 hour forecasts at kilometer scale. It uses generative techniques to directly predict satellite and radar images, simulating storm dynamics. In tests, it outperforms traditional physical models for short-term precipitation. It’s trained on geostationary observations (GOES) for CONUS, but the methodology can be replicated in other regions with similar satellite coverage.
Earth-2 Medium Range (Atlas): covers up to 15 days and predicts 70+ variables (temperature, pressure, wind, humidity, etc.). It uses a latent diffusion transformer architecture that models incremental changes in the atmosphere, which helps preserve critical structures and reduce errors at medium horizons. In standard benchmarks, it outperforms leading open models like GenCast on frequently used variables.
Earth-2 Global Data Assimilation (HealDA, coming soon): generates initial atmospheric states (temperature, wind, humidity, pressure) at thousands of locations in seconds using GPUs, instead of hours on supercomputers. Paired with Atlas, it promises a fully AI-based chain to produce high-quality open forecasts.
These models join NVIDIA’s family of open projects like FourcastNet3, CorrDiff, cBottle and DLESym.
How the tools fit together: Earth2Studio and Physics Nemo
Earth2Studio: an open source Python ecosystem to assemble inference pipelines with the checkpoints available on Hugging Face. Ideal for rapid prototyping and deployment.
Physics Nemo: the stack for training models when you need to tweak architectures or train from your own data. If you want to adapt StormScope to another region or improve Atlas with a local dataset, this is where you do it.
Think of it as a complete toolbox: you train with Physics Nemo, produce initial conditions with HealDA, run medium-range forecasts with Atlas and get detailed nowcasts with StormScope.
What this means for developers and policymakers
For developers: access to modern architectures (generative and diffusion-transformer based) and production-ready checkpoints. You can experiment, fine-tune and deploy on commercial GPUs.
For agencies and governments: the possibility to operate your own models, avoiding exclusive dependence on closed products or third parties. That reduces dependency risks and lets you prioritize local needs, from early warnings for extreme rainfall to agricultural planning.
For the scientific community: a reproducible ecosystem that accelerates iteration and comparison between methods.
Technical considerations and current limitations
Training coverage: StormScope was trained with GOES over CONUS. For regions without similar geostationary satellites, you’ll need to gather equivalent data or adapt the training.
Compute requirements: while HealDA reduces data assimilation times to seconds on GPU, training and tuning these models still requires powerful GPUs and good-quality data.
Operational validation: beating benchmarks is a big step, but integrating into operational chains (real-time observations, data QC, alert communications) requires additional work.
A practical pipeline example
HealDA generates global initial conditions in seconds.
Atlas produces medium-range forecasts (up to 15 days) for global variables.
StormScope takes model outputs plus satellite/radar data to generate kilometer-scale nowcasts for 0–6 hours.
Earth2Studio orchestrates inference and produces outputs ready for visualization or ingestion into alert systems.
What is it good for in practice? Earlier alerts for urban floods, better crop planning, and more precise storm warnings that save lives.
Final reflection
NVIDIA Earth-2 is not just a model release. It’s a bet on opening the full prediction chain to the community, combining generative techniques, diffusion transformers and GPU-accelerated pipelines. If you work with weather or environmental data, this toolbox gives you a powerful starting point to build local, reproducible capabilities.
