Hugging Face and Earthmover published a Sept. 8, 2026 joint guide showing open artificial intelligence weather models still need initialization data, storage, bandwidth, and compatible hardware.
The post says open weights alone are not enough. It cites compute limits such as flash-attention and GPU-architecture requirements for models including ECMWF AIFS, roughly 360 GB of disk for a full-year backtest before outputs, and about 1 GB of initial conditions per typical forecast while inference itself takes seconds.
A Hugging Face demo Space combines Earthmover initialization data with Hub weather models, compares forecasts with ERA5, and reports an end-to-end 24-hour forecast in under about 30 seconds. That demo timing is separate from the local Aurora tutorial, which the post says needs no GPU, takes 2 to 3 minutes per step on CPU, a few seconds on GPU, and does not support Apple MPS.
The local walkthrough uses Microsoft's Aurora 0.25° pretrained checkpoint with ERA5 from the Earthmover Marketplace for a four-step, 24-hour forecast, one six-hour step per forward pass. The post says those Aurora weights are MIT-licensed on the Hub. For remote runs, it presents Hugging Face Jobs with an NVIDIA A100 Large example priced at $2.50 per hour and a 28-step, seven-day Aurora forecast written to a Hub storage bucket. It attributes data-access gains to Zarr chunking and Icechunk's transactional, versioned layer on top of Zarr.
ECMWF's Winter 2026 newsletter independently says AIFS Single and AIFS ENS weights are openly available on Hugging Face and that its experimental ai-models package includes GraphCast, Pangu-Weather, Aurora, and FourCastNet for comparison with AIFS. Microsoft's Aurora documentation independently identifies Aurora 0.25° Pretrained for ERA5 at 0.25° and shows a local ERA5 rollout example.
What remains open is how the demo's under-30-second path transfers to other hardware, whether planned transformers library integrations ship as described, and how Hub licenses vary across AIFS, Aurora, Atlas, FourCastNet, and Prithvi. Figures for timing, storage, bandwidth, and A100 pricing come from the Hugging Face and Earthmover post and should be weighed separately from ECMWF and Microsoft documentation until independently measured.