Model weights

Each row is the current published build of one reference: the archive, the commit that produced it, its size, the SHA-256 of the bytes you will receive, the two licences that build ships under, and the reference atlas report describing what it was trained on. What each model was built from is tabulated on the methods page.

Commercial use of the weights requires a separate licence: contact us.

Batch labels are published as opaque indices. Model behaviour is unchanged. The full explanation ships as NOTICE.txt inside the archive.

Weights alone reproduce the cell type calls. They do not reproduce the OOD score or the secondary label columns, which are computed against reference artifacts rather than the model: the aggregate posterior and the reference latent space.

Licensing

Each archive carries two licences because it holds two kinds of content: the model parameters, our own work, under a licence that permits research and other non-commercial use, named in the row; and the metadata alongside them (label registry, manifest, dataset list), derived from the training data and passed on under the training data's own terms, with attribution. Both are named in the row, and the archive's NOTICE.txt states which files fall under which. The training data is CZI CELLxGENE Census, CC BY 4.0, attributed per study on the attribution page and in each build's manifest. We do not redistribute reference expression data. See terms.