On January 26, 2026, at the American Meteorological Society’s Annual Meeting, NVIDIA announced Earth-2 — a comprehensive family of open-source AI models, libraries, and frameworks designed to democratize weather and climate forecasting.
Described as the world's first fully open, accelerated weather AI software stack, Earth-2 provides an end-to-end pipeline from processing observational data to generating 15-day global forecasts or hyper-local storm predictions.
By making production-grade AI weather tools freely available, NVIDIA aims to accelerate innovation, reduce costs, and enable more organizations—from national meteorological agencies to startups and researchers — to run sophisticated forecasting systems on their own infrastructure.
What is Earth-2?
Earth-2 is not a single model but an integrated ecosystem that includes:
- Pretrained, state-of-the-art AI models;
- Inference and training frameworks;
- Customization recipes and libraries;
- Tools for data assimilation, forecasting, nowcasting, downscaling, and visualization.
The platform supports every stage of the forecasting workflow: assimilating real-time observations (satellites, radars, weather stations), generating initial atmospheric states, producing medium-range global predictions, and refining them to high-resolution local forecasts. It integrates open models from ECMWF, Microsoft, Google, and others, while NVIDIA contributes its own breakthroughs.
Key enabling technologies include NVIDIA's PhysicsNeMo (an open Python framework for training physics-AI models) and Earth2Studio (for building and running inference pipelines).
Core Models in the Earth-2 Family
1. Earth-2 Medium Range (powered by Atlas architecture)
- Delivers probabilistic 15-day global forecasts across 70+ atmospheric variables (temperature, wind, pressure, humidity, precipitation, etc.).
- Outperforms leading open models on standard forecasting benchmarks.
- Enables ensemble-style predictions for uncertainty quantification.
2. Earth-2 Nowcasting (powered by StormScope)
- Generative AI model for 0–6 hour high-resolution (kilometer-scale) forecasts of severe weather, storms, precipitation, and radar/satellite imagery.
- First open AI model to outperform traditional physics-based models on short-term precipitation accuracy.
- Learns to simulate realistic storm evolution and organization directly from observations.
3. Earth-2 Global Data Assimilation (powered by HealDA)
- Produces high-quality initial atmospheric conditions (temperature, wind, humidity, pressure) in seconds on GPUs — instead of hours on supercomputers.
- When coupled with Earth-2 Medium Range, forms the most skillful fully open AI forecasting pipeline to date.
4. Supporting Models
- Earth-2 CorrDiff: Generative downscaling architecture that transforms coarse global predictions into high-resolution (e.g., 2.5 km) regional fields up to 500x faster than traditional methods, with 10,000x better energy efficiency.
- Earth-2 FourCastNet3: High-accuracy global forecasting model for variables like wind and temperature; up to 60x faster than diffusion-based alternatives while rivaling top-tier accuracy.
Why This Matters: Democratizing Weather Intelligence
Traditional numerical weather prediction (NWP) relies on physics-based simulations that demand massive supercomputing resources—limiting access to wealthy nations and large institutions.
Earth-2 flips this paradigm:
- Speed & Cost: AI models run orders of magnitude faster on NVIDIA GPUs, slashing compute time and energy use.
- Accuracy: In several benchmarks (especially short-term precipitation and downscaling), Earth-2 models already match or exceed physics-based systems.
- Openness: All models, weights, code, and documentation are available on Hugging Face (collections/nvidia/earth-2), GitHub (NVIDIA/earth2studio), and NVIDIA's developer resources — under permissive licenses.
- Customization: Developers can fine-tune models on local data, adapt for specific regions, or build sovereign forecasting systems without vendor lock-in.
Real-world early adopters already demonstrate impact:
- Israel Meteorological Service: Uses CorrDiff for operational high-resolution forecasts with 90% compute reduction.
- Brightband: Runs Earth-2 Medium Range operationally for global predictions.
- TotalEnergies, Eni, S&P Global, AXA: Leverage models for energy risk, grid operations, and insurance scenario modeling.
Also read:
- GameStop's First Swallow: Early Cracks Hit Strategy's Bitcoin Empire
- Anthropic Unveils Claude's New Constitution: A Blueprint for AI Alignment
- The Precision of Automation: How China’s Rifle-Wielding Drones Are Redefining the Modern Battlefield
The Bigger Picture: Toward Climate Resilience
Accurate, timely weather and climate information saves lives, protects economies, and informs decisions in agriculture, energy, disaster response, and public health. As extreme events intensify due to climate change, faster and more affordable forecasting becomes critical.
NVIDIA's vision with Earth-2 is to shift weather AI from an elite, compute-intensive discipline to a widely accessible capability. By open-sourcing the full stack, the company invites global collaboration to push boundaries further — whether improving extreme event prediction, integrating new data sources, or advancing long-term climate modeling.
As Mike Pritchard, NVIDIA's Director of Climate Simulation Research, noted: once trained, these AI models are "1,000 times faster" than conventional methods— unlocking new possibilities for science, business, and society.
Earth-2 is now available for anyone to download, run, fine-tune, and deploy. The future of weather forecasting just became dramatically more open—and dramatically faster.
