{"architecture":{"backbone":"ViT-H/14 (Virchow2G: ViT-G, 1.9B)","family":"Vision Transformer","input":"224\u00d7224 tiles at 5\u00d7, 10\u00d7, 20\u00d7 or 40\u00d7","output":"2,560-dim embedding","params":"632M","pretraining":"DINOv2 with mixed-magnification tiling and domain-specific augmentation"},"article":null,"cancer_slugs":["pan-cancer"],"category":"pathology","confidence":"medium","datasets":[],"developer":"Paige / Memorial Sloan Kettering Cancer Center","evaluation":[{"benchmark":"12 zadan poziomu kafelka (tile-level) w patomorfologii obliczeniowej","external":false,"metric":"porownanie z najlepszymi konkurencyjnymi modelami","source":"https://arxiv.org/abs/2408.00738","value":"wynik na poziomie stanu wiedzy (state of the art) na wszystkich 12 zadaniach w chwili publikacji; autorzy nie podaja w streszczeniu wartosci metryk ani nazw poszczegolnych zadan"},{"benchmark":"Skala modelu i pretreningu (dane deklarowane przez autorow)","external":false,"metric":"liczba parametrow / rozmiar zbioru pretreningowego","source":"https://arxiv.org/abs/2408.00738","value":"Virchow2: 632 mln parametrow (vision transformer); pretrening na 3,1 mln preparatow histopatologicznych. W tej samej pracy opisano rowniez Virchow2G (1,9 mld parametrow) i Virchow2G Mini (22 mln, destylacja Virchow2G) \u2014 sa to ODREBNE modele, nie warianty tej karty"}],"hf":{"downloads":110111,"fetched_at":"2026-09-09T21:33:08Z","gated":"auto","last_modified":"2024-10-26","library":"timm","license":"cc-by-nc-nd-4.0","likes":151,"pipeline_tag":"image-feature-extraction"},"kind":"foundation","license":"cc-by-nc-nd-4.0","limitations":"- Preprint-level evaluation at release; independent benchmarks are still accumulating.\n- Non-commercial licence.","links":{"demo":null,"docs":null,"doi":null,"github":null,"huggingface":"https://huggingface.co/paige-ai/Virchow2","paper":"https://arxiv.org/abs/2408.00738","pmid":null},"modalities":["histopathology"],"name":"Virchow2","notable_uses":"","openness":"gated-weights","regulatory":{"intended_use_en":"Research release.","intended_use_pl":"Wydanie badawcze.","source_url":"https://huggingface.co/paige-ai/Virchow2","status":"research-only"},"regulatory_status":"research-only","release_date":"2024-08-01","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:paige-ai/Virchow2', 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"],"slug":"virchow2","sources":[{"label":"Zimmermann E et al. Virchow2: Scaling Self-Supervised Mixed Magnification Models in Pathology. arXiv 2024","url":"https://arxiv.org/abs/2408.00738"},{"label":"Hugging Face \u2014 paige-ai/Virchow2","url":"https://huggingface.co/paige-ai/Virchow2"}],"summary":"Successor to Virchow: ViT-H/14 pretrained on 3.1 million whole-slide images from about 225,000 patients across 45 countries, at mixed magnifications (5\u00d7\u201340\u00d7), with pathology-specific augmentations.","tasks":["feature-extraction","classification","detection"],"training":{"institutions":"MSK and partner institutions worldwide","populations":"multi-national","size":"3.1M WSIs","summary_en":"3.1 million WSIs, ~225,000 patients, 45 countries; H&E plus a share of IHC; the largest multi-site pathology pretraining set reported at release.","summary_pl":"3,1 mln preparat\u00f3w, ok. 225 000 pacjent\u00f3w, 45 kraj\u00f3w; H&E i cz\u0119\u015b\u0107 IHC; najwi\u0119kszy wieloo\u015brodkowy zbi\u00f3r pretreningowy w patomorfologii w chwili wydania."},"updated_at":"2026-09-09T21:33:09.703723","url":"/ai-oncology/models/virchow2","usage":{"library":"timm","notes_en":"Gated on Hugging Face; non-commercial. Same embedding recipe as Virchow (class token + mean patch tokens)."},"verified_at":"2026-09-05T22:26:01.142049","verified_by":"editorial","version":null,"what_it_does":"Tile embeddings robust across magnifications and scanners, intended as a drop-in improvement over Virchow for every downstream slide-level task."}