scGPT

weights

open(ungated whole-human, organ-specific, and pan-cancer checkpoint folders)

data

closed(33 million pretraining cells described in the preprint with no corpus or reconstruction scripts released)

code

partial(fine-tuning, inference, and reference-mapping examples

license

MIT(OSI)

scGPT is an openly licensed checkpoint zoo with adaptation code, but the repository's own to-do list still carries the pretraining code, and the corpus was never released.

• https://raw.githubusercontent.com/bowang-lab/scGPT/main/README.md recorded 2026-09-10

  Lists eight downloadable pretrained checkpoint folders, describes pretraining on 33 million normal human cells without releasing the corpus, ships a fine-tuning example, and leaves "Provide the pretraining code with generative attention masking" unchecked.

• https://raw.githubusercontent.com/bowang-lab/scGPT/main/LICENSE recorded 2026-09-10

  MIT License body covering the repository, read directly rather than from the license endpoint.

No Hugging Face artifact exists - publishing to the Hub is still an open to-do in the repository - so 1,626 GitHub stars are the last-resort signal. One channel is undeclared and therefore not banded on: the scgpt PyPI package draws 3,190 downloads in the trailing 30 days, which would put it at level 1 rather than the 2 the stars produce. Declaring it is a maintainer call, and it would lower this band rather than raise it.

• https://api.github.com/repos/bowang-lab/scGPT recorded 2026-09-10

  stargazers_count: 1626; archived: false; fork: false.

Three attributes are established. The multi-omic input is not merely documented - tutorials/Tutorial_Multiomics.ipynb ships in the repository at about 2.1 MB, alongside Tutorial_Perturbation.ipynb, so the unchecked to-do line is stale relative to what the tree contains. Gene and cell representations are treated as one scale, and the organ-specific checkpoints are separate models rather than a second scale.

• https://raw.githubusercontent.com/bowang-lab/scGPT/main/README.md recorded 2026-09-10

  Ships examples/finetune_integration.py for fine-tuning a pretrained checkpoint and names the model a generative pre-trained foundation model for single-cell multi-omics.

• https://scgpt.readthedocs.io/en/latest/introduction.html recorded 2026-09-10

  Describes generative pre-training on over 33 million cells, single-cell multi-omic data analysis, and downstream multi-omic integration and genetic perturbation prediction, with fine-tuning tutorials for each.