Stanford SNAP

weights

open(ungated 4-layer and 33-layer checkpoints, fetched by the evaluation script or from Figshare)

data

documented-not-released(the paper describes the 36-million-cell Integrated Mega-scale Atlas corpus while the repository links processed paper datasets rather than the pretraining corpus)

code

partial(embedding and evaluation script only

license

CC-BY-4.0(governs the Figshare-hosted checkpoints

The checkpoints are downloadable and the paper describes the training corpus, but the repository ships only code to embed a dataset, not a pretraining pipeline, keeping this off a fully open release. The weights' own license applies, not the repository's: the 33-layer checkpoint ships from a Figshare record stating CC BY 4.0, while the repository LICENSE is MIT and covers only the code.

• https://raw.githubusercontent.com/snap-stanford/UCE/main/LICENSE recorded 2026-09-10

  MIT License body covering the repository and the released model files, read directly rather than from the license endpoint.

• https://raw.githubusercontent.com/snap-stanford/UCE/main/README.md recorded 2026-09-10

  Documents automatic download of the pretrained 4-layer model and a Figshare link for the 33-layer model, exposes only eval_single_anndata.py, and links processed paper datasets rather than the training corpus.

• https://www.nature.com/articles/s41586-026-10689-z recorded 2026-09-10

  Describes self-supervised training on a large corpus of cell data and the resulting Integrated Mega-scale Atlas of 36 million cells from hundreds of experiments, dozens of tissues, and eight species.

No Hugging Face or package artifact is declared, so GitHub stars are used as the last-resort signal; the checkpoints are served from the authors' own hosting and from Figshare, neither of which reports a download count.

• https://api.github.com/repos/snap-stanford/UCE recorded 2026-09-10

  stargazers_count: 338; archived: false; fork: false.

The absent adaptation pathway is a positive claim by the publisher, not a gap in the evidence: UCE is built to embed new cells with no additional training or fine-tuning. It also emits embeddings rather than generated outputs, takes one transcriptomic modality, and works at one scale -- embedding cells from many species broadens coverage of that same modality rather than adding a second one.

• https://raw.githubusercontent.com/snap-stanford/UCE/main/README.md recorded 2026-09-10

  Exposes a single embedding script whose output is UCE embeddings added to the input AnnData, with no training, fine-tuning, or generation entry point.

• https://www.nature.com/articles/s41586-026-10689-z recorded 2026-09-10

  States that UCE's universality means new cells can be embedded with no data labelling, model training or fine-tuning, and describes the output as a unified biological latent space.