Nvidia open-sources Earth-2 AI weather models
Nvidia just launched the Earth-2 family of open models, delivering the world's first fully open, accelerated software stack for AI weather and climate forecasting. The suite lets organizations run and fine-tune models on their own hardware, cutting compute time by up to 90% in operational tests by meteorological agencies. Why it matters: It shifts heavy supercomputer weather forecasting into accessible AI workflows that any enterprise or government can run locally.
Nvidia's Earth-2 is an open-source suite of AI models and accelerated software built to handle end-to-end weather and climate forecasting on your own hardware. Instead of relying on supercomputer-heavy physics equations, it packages everything from data processing to 15-day global outlooks and local storm tracking into an accessible, developer-friendly stack.
What is inside and how it works
Traditional numerical weather prediction eats massive supercomputing time solving fluid dynamics equations. Earth-2 replaces or augments parts of this pipeline with physical AI, bundling pretrained models, customization recipes, and inference acceleration libraries.
Key additions to the stack include Earth-2 Medium Range for medium-term forecasting and Earth-2 Nowcasting for high-resolution short-term severe weather. They sit alongside existing open models like the downscaling tool CorrDiff and the global forecast model FourCastNet3. A dedicated Global Data Assimilation module is slated to drop later this year.
The framework also integrates external open models from ECMWF, Google, and Microsoft. Developers looking to fine-tune these architectures on regional atmospheric data can use PhysicsNeMo, Nvidia's open-source Python framework designed for training physics-informed AI at scale.
The numbers and real-world results
The speedups reported in operational trials are substantial. The Israel Meteorological Service put Earth-2 CorrDiff to work at a 2.5-kilometer resolution and saw a 90% reduction in compute time compared to running conventional numerical prediction on a CPU cluster. During a recent heavy storm, its AI setup outperformed every operational baseline model in six-hour accumulated precipitation accuracy.
Commercial players are already relying on it. Startup Brightband runs Earth-2 Medium Range daily to produce operational global forecasts. Solar operator GCL integrated Earth-2 to improve photovoltaic output predictions at lower compute overhead.
Grid operator Southwest Power Pool and Hitachi are applying Nowcasting and FourCastNet3 to intraday wind forecasting, while energy giants TotalEnergies and Eni test the suite for risk assessment and natural gas demand modeling.
How to get started
Nvidia has open-sourced both the code and model weights for Earth-2 Medium Range and Earth-2 Nowcasting. You can pull them right now from GitHub and Hugging Face, or test them through the Earth2Studio toolkit.
The full suite is still rolling out: the data assimilation system is scheduled for later this year, and while you no longer need a national lab's supercomputer, serious local fine-tuning and high-resolution inference will still demand serious GPU iron.
Why it matters
Weather forecasting has historically been locked behind government-level supercomputing facilities. By delivering an open, production-ready pipeline that organizations can run on their own infrastructure, Nvidia makes high-resolution weather intelligence practical for regional forecasters, renewable energy operators, and insurers.
Strategically, Nvidia is playing the long game with physical AI. Giving away state-of-the-art models establishes its software stack as the default standard for climate tech, while guaranteeing that the massive compute workloads migrate directly to its GPUs. For anyone building in energy trading, grid operations, or disaster risk, this is an immediate green light to start benchmarking against legacy numerical setups.
Sources (1)
- blogs.nvidia.com(opens in a new tab)Originalblogs.nvidia.com