Overview
HyCell is a 16.1-million-parameter single-cell transcriptome embedding model released on November 2, 2026, with weights and code as part of Tencent’s Hyra research artifacts. It was pretrained without cell-type labels on 79.5 million human cells from the CZ CELLxGENE Census. (github.com)
Capabilities and uses
- Maps a human cell’s raw gene counts to a 512-dimensional embedding and can decode embeddings into expected gene expression.
- Supports research workflows including cell clustering, integration across batches, annotation transfer and comparison of cell states. With a labelled reference, it can be fine-tuned for cell-type annotation.
- Tencent reports a mean overall score of 0.735 on seven atlases excluded from pretraining, compared with 0.736 for Stack under the reported evaluation. The benchmark covers cell-level tasks, not sample-level disease-state prediction. (raw.githubusercontent.com)
Availability and limitations
The hycell Python package includes approximately 64 MB of weights and can run offline after installation. It is released under Apache-2.0; the repository also identifies CC BY 4.0 attribution requirements for the cited training and test data. A CUDA GPU is recommended, although CPU inference is supported. (raw.githubusercontent.com)
The model expects raw integer counts and uses a vocabulary of 36,495 human genes; genes outside that vocabulary are dropped. The authors recommend supplying all measured genes rather than only a highly variable subset. It was trained on human primary cells, so performance should not be assumed to generalize to other species or untested datasets. (raw.githubusercontent.com)
Pricing
No official HyCell-specific API input or output pricing could be verified in the mainland China or international Tencent Cloud pricing sources checked on November 14, 2026. The published local package and weights do not establish a hosted API price; infrastructure costs for running them are separate. (cloud.tencent.com)