ECMWF

weights

open(CC-BY-4.0 ungated aifs-single-mse-1.0.ckpt)

data

documented-not-released(ERA5 1979-2022 for pretraining and operational IFS analyses 2016-2022 for fine-tuning, with the full input and output field list

code

partial(config_pretraining.yaml and config_finetuning.yaml ship beside the checkpoint with pinned Anemoi versions and both stage commands, but no data-preparation recipe - the commands begin DATASETS_PATH=??????? and no anemoi-datasets build ships)

license

CC-BY-4.0(permissive non-OSI)

ECMWF publishes both training-stage configs beside the checkpoint - pinned Anemoi versions, learning-rate schedules, step counts and the commands that run them - which is more than most of the category ships. It is recorded as partial rather than open because the rubric's `open` asks for data prep, training recipe and configs, and the data-preparation half does not ship: both commands begin `export DATASETS_PATH=???????` with no anemoi-datasets build to fill it. Against the category's other `open` records - presto's end-to-end data-generation pipeline, mattergen's full preprocessing-to-evaluation pipeline - two YAMLs and a pinned third-party trainer do not reach the same bar.

• https://huggingface.co/ecmwf/aifs-single-1.0/raw/main/README.md recorded 2026-09-10

  "These model weights are published under a Creative Commons Attribution 4.0 International (CC BY 4.0)" while notebooks and scripts are Apache 2.0; gives the pretraining command anemoi-training train --config-name=config_pretraining.yaml against pinned anemoi-training 0.3.1, anemoi-models 0.4.0 and anemoi-graphs 0.4.4, then the fine-tuning stage; and details pretraining on ERA5 1979-2022 followed by rollout on operational IFS analyses 2016-2022.

• https://huggingface.co/api/models/ecmwf/aifs-single-1.0 recorded 2026-09-10

  license: cc-by-4.0; gated: false; private: false; siblings include aifs-single-mse-1.0.ckpt, config_pretraining.yaml and config_finetuning.yaml.