{"architecture":{"backbone":"ViT-g/14 (tile) + LongNet (slide)","family":"ViT tile encoder + LongNet slide encoder","input":"256\u00d7256 tiles at 20\u00d7; slide-level sequence of all tile embeddings with coordinates","notes_en":"The slide encoder is the distinctive part: dilated attention keeps memory manageable for sequences of 10,000+ tiles, so context from across the slide informs each token.","output":"1,536-dimensional tile embeddings; slide-level embedding","params":"1.3B","pretraining":"DINOv2 (tile) + masked autoencoder objective over tile embeddings (slide)"},"article":null,"cancer_slugs":["pan-cancer"],"category":"pathology","confidence":"high","datasets":[{"name":"TCGA \u2014 The Cancer Genome Atlas (via NCI Genomic Data Commons)","note":"","role":"evaluation","slug":"tcga"}],"developer":"Microsoft Research / Providence Health / University of Washington","evaluation":[{"benchmark":"pan-cancer subtyping (9 cancer types, Providence + TCGA)","dataset_slug":"tcga","external":true,"metric":"AUROC / balanced accuracy","source":"https://www.nature.com/articles/s41586-024-07441-w","value":"best in 25/26 tasks vs. HIPT, CtransPath, REMEDIS"},{"benchmark":"driver-mutation prediction from H&E (18 genes, pan-cancer)","external":true,"metric":"AUROC","source":"https://www.nature.com/articles/s41586-024-07441-w","value":"reported per gene in paper"}],"hf":{"downloads":45031,"fetched_at":"2026-09-09T21:33:08Z","gated":"auto","last_modified":"2026-08-07","library":"timm","license":"apache-2.0","likes":189,"pipeline_tag":"image-feature-extraction"},"kind":"foundation","license":"apache-2.0","limitations":"- Single health system: scanner and staining conventions of Providence dominate the pretraining distribution.\n- Slide-level encoder is memory-hungry; small labs may only use the tile encoder.\n- Mutation predictions from H&E are associative, not a substitute for sequencing.","links":{"demo":null,"docs":null,"doi":"10.1038/s41586-024-07441-w","github":"https://github.com/prov-gigapath/prov-gigapath","huggingface":"https://huggingface.co/prov-gigapath/prov-gigapath","paper":"https://www.nature.com/articles/s41586-024-07441-w","pmid":null},"modalities":["histopathology","multimodal"],"name":"Prov-GigaPath","notable_uses":"State of the art on 25 of 26 tasks at publication, including pan-cancer subtyping and prediction of common driver mutations from H&E; the first pathology model to be trained end-to-end at slide scale on a health-system archive.","openness":"gated-weights","regulatory":{"intended_use_en":"Research; explicitly not for clinical use per the repository.","intended_use_pl":"Badania; wed\u0142ug repozytorium wyra\u017anie nie do u\u017cytku klinicznego.","source_url":"https://github.com/prov-gigapath/prov-gigapath","status":"research-only"},"regulatory_status":"research-only","release_date":"2024-05-22","run_snippet":"# generic timm loader \u2014 check the model card for the exact init args and image normalisation\nimport timm, torch\nfrom huggingface_hub import login\nlogin()  # gated repos: accept the licence on huggingface.co first\nmodel = timm.create_model('hf-hub:prov-gigapath/prov-gigapath', pretrained=True)\nmodel.eval()\ncfg = timm.data.resolve_data_config({}, model=model)\ntransform = timm.data.create_transform(**cfg)\n# emb = model(transform(tile).unsqueeze(0))  # 1 x D tile embedding\n","settings":["basic-research","diagnosis","prognosis"],"slug":"prov-gigapath","sources":[{"label":"Xu H et al. A whole-slide foundation model for digital pathology from real-world data. Nature 2024","url":"https://www.nature.com/articles/s41586-024-07441-w"},{"label":"Hugging Face \u2014 prov-gigapath/prov-gigapath","url":"https://huggingface.co/prov-gigapath/prov-gigapath"}],"summary":"Whole-slide foundation model with 1.3 billion parameters, pretrained on 1.3 billion tiles from 171,189 slides of real-world clinical data; pairs a DINOv2 tile encoder with a LongNet slide encoder that reasons over an entire slide.","tasks":["feature-extraction","classification","prognosis","image-text-retrieval"],"training":{"consent_notes":"De-identified institutional data; the archive is not released \u2014 only the weights.","institutions":"Providence Health & Services (28 cancer centres, USA)","populations":"US community and academic oncology population; real-world case mix","size":"171,189 WSIs / 1.3B tiles","summary_en":"Providence health system archive: 171,189 H&E and IHC whole-slide images from more than 30,000 patients across 31 tissue types, covering 28 cancer centres in the US; 1.3 billion tiles at 256\u00d7256.","summary_pl":"Archiwum systemu Providence: 171 189 preparat\u00f3w H&E i IHC od ponad 30 000 pacjent\u00f3w, 31 typ\u00f3w tkanek, 28 o\u015brodk\u00f3w onkologicznych w USA; 1,3 mld kafelk\u00f3w 256\u00d7256."},"updated_at":"2026-09-09T21:33:09.703722","url":"/ai-oncology/models/prov-gigapath","usage":{"hardware":"tile encoder: 1 GPU; slide encoder over a full slide: \u226524 GB GPU memory recommended","library":"timm + repo code (slide encoder)","notes_en":"Weights are Apache-2.0 but gated (accept terms on Hugging Face). The repo ships the tile\u2192slide pipeline; keep tile coordinates, the slide encoder needs them."},"verified_at":"2026-09-05T22:26:01.122323","verified_by":"editorial","version":null,"what_it_does":"Two stages: a tile encoder produces embeddings, then a slide encoder built on LongNet (dilated attention) aggregates tens of thousands of tile embeddings into a slide representation. That slide vector feeds cancer subtyping, mutation prediction and vision-language tasks."}