{"architecture":{"family":"Gemma 3 decoder LLM + SigLIP image encoder (MedSigLIP)","input":"images + text (instruction format)","output":"text","params":"4B (multimodal) / 27B (text)","pretraining":"Gemma 3 pretraining + medical image\u2013text and text tuning"},"article":null,"cancer_slugs":["pan-cancer"],"category":"clinical-LLM","confidence":"medium","datasets":[{"name":"MedQA (USMLE)","note":"","role":"benchmark","slug":"medqa"}],"developer":"Google (Health AI Developer Foundations)","evaluation":[{"benchmark":"MedQA (USMLE-style, text)","dataset_slug":"medqa","external":true,"metric":"accuracy","source":"https://arxiv.org/abs/2507.05201","value":"reported in technical report (27B variant)"}],"hf":{"downloads":1105085,"fetched_at":"2026-09-09T21:33:09Z","gated":"auto","last_modified":"2025-10-28","library":"transformers","license":"other","likes":1048,"pipeline_tag":"image-text-to-text"},"kind":"foundation","license":"Health AI Developer Foundations terms of use","limitations":"- Generative: can produce fluent but wrong findings; needs task-specific evaluation and guardrails.\n- Oncology-specific validation is limited in the report; pathology capability is preliminary.","links":{"demo":null,"docs":"https://developers.google.com/health-ai-developer-foundations/medgemma","doi":null,"github":null,"huggingface":"https://huggingface.co/google/medgemma-4b-it","paper":"https://arxiv.org/abs/2507.05201","pmid":null},"modalities":["multimodal","radiology-xray","histopathology","dermoscopy","clinical-text"],"name":"MedGemma","notable_uses":"","openness":"gated-weights","regulatory":{"intended_use_en":"Developer foundation for building and validating your own medical applications; not a medical device.","intended_use_pl":"Baza deweloperska do budowy i walidacji w\u0142asnych aplikacji medycznych; nie jest wyrobem medycznym.","source_url":"https://developers.google.com/health-ai-developer-foundations/medgemma","status":"research-only"},"regulatory_status":"research-only","release_date":"2025-05-20","run_snippet":"# generic transformers loader \u2014 see the model card for the task-specific head and preprocessing\nfrom transformers import AutoModel, AutoProcessor\nmodel = AutoModel.from_pretrained('google/medgemma-4b-it')\nprocessor = AutoProcessor.from_pretrained('google/medgemma-4b-it')\n","settings":["basic-research","education"],"slug":"medgemma","sources":[{"label":"MedGemma Technical Report (arXiv 2507.05201)","url":"https://arxiv.org/abs/2507.05201"},{"label":"Health AI Developer Foundations \u2014 MedGemma","url":"https://developers.google.com/health-ai-developer-foundations/medgemma"},{"label":"Hugging Face \u2014 google/medgemma-4b-it","url":"https://huggingface.co/google/medgemma-4b-it"}],"summary":"Open-weight medical vision-language models built on Gemma 3: the 4B variant reads chest X-rays, dermatology, ophthalmology and histopathology images alongside text; the 27B variant targets medical text reasoning. Meant as a starting point for developers to fine-tune, not as a finished clinical product.","tasks":["question-answering","report-generation","classification","information-extraction"],"training":{"summary_en":"Gemma 3 base plus medical fine-tuning on de-identified image\u2013text data (chest X-ray, dermatology, ophthalmology, histopathology) and medical text; details in the technical report.","summary_pl":"Baza Gemma 3 plus dostrajanie medyczne na zanonimizowanych danych obraz\u2013tekst (RTG klatki piersiowej, dermatologia, okulistyka, histopatologia) i tekstach medycznych; szczeg\u00f3\u0142y w raporcie technicznym."},"updated_at":"2026-09-09T21:33:09.703717","url":"/ai-oncology/models/medgemma","usage":{"hardware":"4B runs on a single 16\u201324 GB GPU; 27B needs multi-GPU or quantisation","library":"transformers","notes_en":"Accept the HAI-DEF terms on Hugging Face. Outputs must be validated per task; Google states the models are not intended for direct clinical use without further evaluation."},"verified_at":"2026-09-05T22:26:00.309905","verified_by":"editorial","version":"4B multimodal / 27B text","what_it_does":""}