Model reference · open weights

UNI2-h

UNI2-h is an open-weight embedding model from MahmoodLab. UNI2-h (BF16) weighs 2.7 GB; the smallest configuration that runs it is RTX 3060 12 GB.

• UNI2-h is a pretrained vision backbone developed by MahmoodLab for image feature extraction on histopathology images.
• The model utilizes a custom ViT-H/14 architecture with 681 million parameters and was trained on over 200 million image tiles from H&E and IHC slides.
• It is released under the CC-BY-NC-ND-4.0 license, restricting use to non-commercial academic research purposes.

Summary of the MahmoodLab/UNI2-h model card, 2026-10-01

Hugging Face ↗