The Mahmood Lab's updated tile-level pathology encoders: UNI2-h (a DINOv2 vision FM) and CONCH v1.5 (a vision-language model). These are the field's most-adopted pathology encoders — UNI2-h ~162K, the original UNI ~73K, and CONCH ~83K HuggingFace downloads. The original UNI/CONCH appeared in Nature Medicine 2024; these are the in-window refreshes. Pretrained from scratch on pathology images by Harvard.

CONCH-Z (2026-10-12; arXiv 2610.17233, a NeurIPS 2026 spotlight) is a lineage checkpoint of CONCH v1.5, not a new pretraining run. The CONCH v1.5 ViT-L/16 vision encoder and its contrastive projection stay frozen. Only a new text tower is trained, with LIMIT (Locked Image Multiple Instance PreTraining), against about 335K slide-level clinical pathology reports in place of curated image captions. Report embeddings act as cross-attention queries over bags of patch features and are trained with InfoNCE. The lab's README reports zero-shot gains over CONCH v1.5 of 90.3→96.7 AUC on CAMELYON-16, 49.7→58.8 and 42.9→51.3 balanced accuracy on SPIDER coarse and fine, and 47.9→64.4 on CATCH canine skin. It also beats PLIP, QuiltNet, MUSK and MI-Zero on all four. These are the authors' numbers across the paper's 22 tasks; none is independently measured. The parameter count is not stated. The weights are gated on HuggingFace under CC-BY-NC-ND 4.0, for non-commercial academic use only, and the LIMIT training code is listed as coming soon.

Model Details

sciencebiologyvisionopen-weight