AIFS

The Artificial Intelligence Forecasting System is the European Centre for Medium-Range Weather Forecasts' data-driven weather model. It uses a graph neural network encoder and decoder with a transformer processor, trained first on ERA5 reanalysis for 1979 to 2022 and then fine-tuned on ECMWF's own operational analyses. It runs alongside, not instead of, the physics-based Integrated Forecasting System, which still supplies the initial conditions.

It forecasts upper-air and surface variables, including 100 m winds, solar radiation, cloud cover, snowfall, soil moisture and tropical cyclone tracks, at about 28 to 31 km grid spacing and 13 pressure levels, in 6-hour steps. ECMWF reports that it beats its physics-based models on many scores and uses roughly a thousand times less energy per forecast.

The weights of AIFS Single v1.0 are published on Hugging Face under CC BY 4.0, with Apache-2.0 notebooks for running them. The model was trained with Anemoi, ECMWF's open framework for machine-learned forecasting.

The weights, including the current aifs-single-2.0 checkpoint, are on Hugging Face, but ECMWF's inference tooling depends on flash-attn, which only builds on Ampere-class NVIDIA GPUs. A July 2026 tutorial from Hugging Face and ECMWF adds a compatibility patch so AIFS Single 2.0 runs on any GPU, including Apple MPS, or on CPU, either on your own machine or as a Hugging Face Job on a 16 GB T4. It covers fetching the two consecutive 6-hour analyses the model needs from ECMWF Open Data (free, no account), running a forecast and plotting the fields; a Hugging Face Space runs the same code in the browser on live data.

Operational: AIFS Single since 25 February 2025 and the 51-member AIFS ENS since 1 July 2025, both with ECMWF's round-the-clock support.