Hugging Face and Earthmover simplify open AI weather forecasting

Open-weight weather models from ECMWF, Microsoft, and Google DeepMind are becoming credible complements to physics-based forecasting—but downloading and running them remains an infrastructure problem. A new joint guide from Earthmover and Hugging Face, published September 8, 2026, shows developers how to make that process more practical.

The Hugging Face Blog pairs runnable demonstrations with a local tutorial. It uses Earthmover’s analysis-ready weather and climate data for model inputs and validation, while also showing how Hugging Face Jobs can handle the compute side.

The bottleneck is often the data

Inference itself can take only seconds on a laptop or small cluster, but getting to that point is less straightforward. Models such as ECMWF’s AIFS may require Flash Attention, which limits compatible GPU hardware. A single forecast typically needs about 1GB of initial-condition data, meaning the accelerator may spend more time waiting for downloads than generating predictions.

Storage becomes a larger issue for experimentation. Backtesting a model across a full year requires roughly 360GB of disk space before forecast outputs are added. That makes repeated evaluation and training difficult for individual researchers and small teams, even when the neural network can run on modest hardware.

The guide’s practical importance is that it treats the full workflow—not just model inference—as the product. Developers can explore AI forecasts in a demo Space, work with standardized scientific data, and learn how to connect open model weights to historical datasets for validation.

For builders working on climate applications, this could reduce the time spent assembling data pipelines and managing transfers, leaving more effort for benchmarking models, testing new conditions, and building useful forecasting tools.

Source: Hugging Face Blog