Summary: On January 26, 2026, NVIDIA unveiled the Earth-2 family of open models at the American Meteorological Society annual meeting, releasing observation processing, 15-day forecasting, and kilometer-scale nowcasting as an open accelerated weather AI stack. The first release includes Earth-2 Medium Range and Nowcasting on GitHub and Hugging Face; Global Data Assimilation uses HealDA and is expected later in 2026. Production users named by NVIDIA include Brightband and the Israel Meteorological Service, alongside evaluating weather agencies and energy, insurance, and finance organizations.
NVIDIA picked the American Meteorological Society annual meeting in Houston, on January 26, 2026, to unveil the Earth-2 family of open models — three new AI weather models that the company bills as the world’s first fully open, accelerated weather AI software stack. For the first time, the whole forecasting pipeline — observation data processing, 15-day global forecasting, kilometer-scale short-range nowcasting — has been put out in open source.
This goes beyond a lab demonstration. Alongside the models, NVIDIA published a list of organizations already running them in production — weather services, energy companies, insurers — showing that AI weather forecasting has moved from papers onto operational duty. Developers gain too: with weather models they can download, fine-tune, and self-host, a new application space opens up.
Atlas, StormScope, HealDA: What Each New Model Does
- Earth-2 Medium Range is built on Atlas, a new architecture that looks up to 15 days ahead and covers more than 70 weather variables (temperature, pressure, wind, humidity). On standard benchmarks for common forecasting variables, NVIDIA says it beats the leading open models.
- Earth-2 Nowcasting is built on the StormScope architecture, a generative model that turns out zero- to six-hour, kilometer-resolution local storm forecasts within minutes by simulating storm dynamics directly and predicting satellite and radar imagery. By NVIDIA’s description, it is the first AI model to beat physics-based models on short-term precipitation forecasting.
- Earth-2 Global Data Assimilation relies on HealDA to generate initial conditions — temperature, wind, humidity, and pressure at thousands of locations worldwide. A GPU finishes the job in seconds where a supercomputer needs hours. This model has not shipped yet and is expected later in 2026.
The Whole Pipeline, Out in the Open: From Initial Conditions to Day 15
Classic numerical weather prediction leans on supercomputers grinding through physics simulations — expensive to run and slow to iterate. Earth-2 cuts the pipeline into open components you can swap instead: HealDA produces the initial conditions, Atlas carries the medium range, StormScope covers the near term, and the existing CorrDiff (generative downscaling, up to 500x faster than traditional methods) and FourCastNet3 (up to 60x faster than conventional ensemble approaches) slot in alongside. Open models from ECMWF, Microsoft, and Google are part of the stack as well. By NVIDIA’s claim, the HealDA-plus-Atlas combination yields the most skillful predictions ever produced by an open, entirely AI pipeline.
The release covers pretrained models, frameworks, customization recipes, and inference libraries, all runnable, fine-tunable, and deployable on your own infrastructure. Medium Range and Nowcasting are already up on GitHub and Hugging Face through Earth2Studio, with training and fine-tuning handled by the open-source Python framework PhysicsNeMo.
Already Live in Production
A production roster tells you more than a model card:
- Brightband, a startup, has moved Medium Range into operational use and generates forecasts every day.
- The Israel Meteorological Service runs CorrDiff in operations — as many as 8 forecasts a day, with compute time cut 90% at 2.5-kilometer resolution — and it came out on top in a six-hour precipitation check after a rainstorm.
- The partner list also names the U.S. National Weather Service and The Weather Company; Taiwan’s Central Weather Administration is evaluating.
- On the energy and finance side: TotalEnergies is evaluating Nowcasting, GCL’s photovoltaic forecasting is live, Southwest Power Pool is working with Hitachi on wind forecasting, and AXA applies FourCastNet to hurricane scenario analysis.
The Case for Opening All of It
Weather forecasting is public-interest infrastructure in the classic sense, and national weather services have spent years leaning on a handful of supercomputing centers. Putting the full accelerated stack in the open lets agencies and startups on tight budgets stand up their own forecasting capability on GPU clusters rather than buying answers from the large forecasting centers. For NVIDIA, it is the platform playbook it always runs: hand over the software, and the value lands on GPU demand — the models cost nothing, but the accelerators that run them still sell.
What Builders and Product Teams Should Take Away
Three angles. One, the cost of getting hold of weather intelligence keeps dropping — teams in energy trading, agricultural insurance, and logistics scheduling can fine-tune regional models of their own instead of stopping at API access. Two, “forecasting as software” resets the iteration rhythm: with data assimilation cut from hours to seconds, forecast-driven products can refresh far more often. Three, the open-model race has visibly crossed over from language models into scientific ones — the trend we highlighted in our opening-of-year outlook — and more “foundation model plus open stack” plays in vertical domains reward getting in early.