ECMWF has open-sourced its AI weather forecasting model, AIFS, making it accessible to the public. The model delivers medium-range global forecasts with significantly reduced energy consumption compared to traditional physics-based models. AIFS Single 2.0, the latest version, is now available on Hugging Face, allowing users to run it on any hardware without requiring specific GPU hardware. The release aims to lower barriers to entry for research and practical applications, enabling users to generate forecasts, inspect outputs, and verify results against observations or other models.
The AIFS model uses a graph neural network (GNN) encoder and decoder, along with a sliding-window transformer processor, trained on Copernicus ERA5 reanalysis data. It is implemented in Anemoi, an open-source framework developed by ECMWF and European national meteorological services. The model is available in two configurations: AIFS Single, which produces deterministic forecasts, and AIFS ENS, an ensemble model with 51 members representing forecast uncertainty. Both versions were upgraded to v2 on 12 May 2026, incorporating additional training data and improved features like data-driven wave and snow forecasts.
AIFS is already in operational use at ECMWF, providing input to flood-awareness systems since September 2025. The model has shown improvements in tropical cyclone track prediction, with up to 20% better performance compared to leading physics-based models. Despite its operational relevance, the model's accessibility has been enhanced through tools like Hugging Face Jobs, which allow users to run forecasts without specialized hardware. This development enables a broader range of users, including students and researchers, to engage with and contribute to weather forecasting using open-source technology.
Source: huggingface
