Aurora

license

MIT(OSI)

weights

open(MIT ungated checkpoints)

data

components-listed(public ERA5, CMIP6, HRES, and GFS sources enumerated

code

partial(model and inference implementation without pretraining pipeline)

Aurora releases the multi-task checkpoint family and describes the public training sources, but not the pipeline that produced the pretrained base.

• https://huggingface.co/api/models/microsoft/aurora recorded 2026-09-10

  license: mit; gated: false; checkpoint files for weather, air pollution, ocean waves, and cyclone tracking.

• https://raw.githubusercontent.com/microsoft/aurora/main/README.md recorded 2026-09-10

  Lists four specialized forecast versions, links all training datasets, and documents loading released checkpoints; no pretraining procedure is provided.

The Hugging Face Hub reports zero downloads, so GitHub stars are used as the last-resort reach signal instead.

• https://api.github.com/repos/microsoft/aurora recorded 2026-09-10

  stargazers_count: 1014; archived: false; fork: false.

Aurora ships an executable fine-tuning pathway: docs/finetuning.md provides a Docker image and a fine-tuning loop built on AuroraPretrained. It is not counted as generative, because it predicts a deterministic future state rather than generating one.

• https://raw.githubusercontent.com/microsoft/aurora/main/docs/finetuning.md recorded 2026-09-10

  A dedicated Fine-Tuning page - "if you wish to fine-tune Aurora for a specific application, you should build on the pretrained version", with AuroraPretrained().load_checkpoint(), a Docker image and a fine-tuning loop. This is the executable pathway ruling (b) requires, which the README alone only describes.

• https://raw.githubusercontent.com/microsoft/aurora/main/README.md recorded 2026-09-10

  Describes adaptation with limited task data, predictive outputs, and separate medium- and high-resolution weather versions.