On January 26, 2026, at the American Meteorological Society annual meeting in Houston, NVIDIA launched the Earth-2 family of open models — three new AI weather models that the company calls the world’s first fully open, accelerated weather AI software stack. From processing observation data and generating 15-day global forecasts to kilometer-scale short-term nowcasting, the entire weather forecasting pipeline has been released in open source form for the first time.
This is not just a research demonstration. NVIDIA published a roster of users already running the stack in production — weather services, energy companies, and insurers — evidence that AI weather forecasting has moved from papers into operational duty. For developers, downloadable, fine-tunable, self-hostable weather models also open up new application space.
Three New Models: Atlas, StormScope, and HealDA
- Earth-2 Medium Range uses a new architecture called Atlas, forecasting up to 15 days ahead across more than 70 weather variables (temperature, pressure, wind, humidity). NVIDIA says it outperforms leading open models on standard benchmarks for common forecasting variables.
- Earth-2 Nowcasting uses the StormScope architecture, a generative model that produces zero- to six-hour, kilometer-resolution local storm forecasts in minutes by simulating storm dynamics directly and predicting satellite and radar imagery. NVIDIA calls it the first AI model to beat physics-based models on short-term precipitation forecasting.
- Earth-2 Global Data Assimilation uses the HealDA architecture to produce initial conditions — temperature, wind, humidity, and pressure at thousands of global locations — in seconds on GPUs versus hours on supercomputers. This one is not yet released and is expected later in 2026.
Opening the Whole Pipeline: Assimilation to 15-Day Forecast
Traditional numerical weather prediction depends on supercomputers grinding through physics simulations — expensive and slow to iterate. Earth-2’s strategy is to split the pipeline into swappable open components: HealDA generates initial conditions, Atlas handles the medium range, StormScope covers short-term nowcasting, joined by the existing CorrDiff (generative downscaling, up to 500x faster than traditional methods) and FourCastNet3 (up to 60x faster than conventional ensemble approaches). The stack also integrates open models from ECMWF, Microsoft, and Google. NVIDIA claims the HealDA-plus-Atlas pairing yields the most skillful predictions ever produced by an open, entirely AI pipeline.
The release includes pretrained models, frameworks, customization recipes, and inference libraries, all runnable, fine-tunable, and deployable on your own infrastructure. Medium Range and Nowcasting are available now through Earth2Studio on GitHub and Hugging Face, with training and fine-tuning via the open-source Python framework PhysicsNeMo.
Who Is Already Running It
Production adoption says more than a model card:
- Startup Brightband runs Medium Range operationally, producing daily forecasts.
- The Israel Meteorological Service uses CorrDiff in operation, running up to 8 forecasts a day with a 90% reduction in compute time at 2.5-kilometer resolution — and it ranked as the best model in a post-rainstorm six-hour precipitation check.
- The U.S. National Weather Service and The Weather Company are on the partner list; Taiwan’s Central Weather Administration is evaluating.
- Energy and finance: TotalEnergies is evaluating Nowcasting, GCL’s photovoltaic forecasting is live, Southwest Power Pool works with Hitachi on wind forecasting, and AXA uses FourCastNet for hurricane scenario analysis.
Why Open-Sourcing Matters
Weather forecasting is classic public-interest infrastructure, and national weather services have long depended on a handful of supercomputing centers. Open-sourcing the full accelerated stack means budget-constrained agencies and startups can build their own forecasting capability on GPU clusters instead of buying answers from large forecasting centers. For NVIDIA, it is the familiar platform play: give the software away, and let the value settle into demand for GPUs — the models are free, but the accelerators that run them still sell.
What It Means for Builders
Three angles. First, the cost of acquiring weather intelligence is collapsing — applications in energy trading, agricultural insurance, and logistics scheduling can fine-tune their own regional models instead of buying API access alone. Second, “forecasting as software” changes the iteration rhythm: data assimilation shrinking from hours to seconds means forecast-driven products can refresh far more often. Third, this is concrete evidence that 2026’s open-model race is spilling over from language models into scientific models — echoing the trend we flagged in our opening-of-year outlook — and more “foundation model plus open stack” combinations in vertical domains are worth positioning for early.
Sources
- NVIDIA Launches Earth-2 Family of Open Models — NVIDIA Blog
- Nvidia launches Earth-2 open AI weather forecast models and tools — SiliconANGLE
- NVIDIA Launches Earth-2 Family of Open Models for Weather and Climate AI — HPCwire
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.