Hugging Face and Earthmover published a joint guide on Sept. 8, 2026. It covers AI weather forecasting models that are open, meaning their trained files are published so anyone can download them and run them on their own computers. Downloading a model is only the first step, the guide says: a forecast also needs starting weather data, room to store it, fast connections and compatible hardware. Those practical hurdles, not the models themselves, are the real obstacle for weather agencies and labs that want to run their own AI forecasts, the companies argue.
The guide lists what gets in the way. Some models, including AIFS from the European Centre for Medium-Range Weather Forecasts (ECMWF), depend on particular chip designs and supporting software to run well. A full year of test forecasts can take hundreds of gigabytes of disk before any results are saved, and each forecast needs a large download of starting conditions, even though the model itself runs in seconds.
To make this easier, Hugging Face built an online demo that pairs Earthmover's starting data with weather models hosted on Hugging Face and lets users check the forecasts against past weather data. The guide says the demo delivers a 24-hour forecast, start to finish, in under about 30 seconds. It credits two storage formats for easing the storage and download load: Zarr, which keeps huge weather datasets as many small pieces that can be read on their own, and Icechunk, which adds version tracking and safe updates on top of Zarr.
A separate tutorial in the guide runs Microsoft's Aurora model on an ordinary computer with no AI chip at all, though each six-hour forecast step then takes minutes rather than seconds. The guide says Aurora is published under the MIT license, a license that lets anyone use, change and share it free of charge. For bigger jobs, it shows how to rent an NVIDIA A100 machine through Hugging Face's Jobs service at $2.50 an hour and walks through a seven-day Aurora forecast on it.
The guide covers several forecasting models anyone can download, including ECMWF's AIFS, Microsoft's Aurora, NVIDIA's Atlas and FourCastNet, and Prithvi. ECMWF says separately that its AIFS models are openly available on Hugging Face, and Microsoft's own documentation shows Aurora running on a local computer.
The speed, storage and price figures all come from Hugging Face and Earthmover. It is not yet clear how the demo's speed carries over to other hardware, or whether planned support for weather models in Hugging Face's transformers software will ship as described. Each model also comes with its own license terms.