The GreyBrain model formulary
You would read the insert
before you prescribe.
Every one of these models comes with a note from the people who built it — who they are, what it learnt from, whether you may use it and where it stops. We read that note and set it out plainly. Nothing is invented, and a field marked not stated means the makers did not say, which is itself worth knowing before you rely on it.
medgemma-4b-it
Google's medical model. It takes an image and a history together, so it is the one people point to when they claim AI can read a scan. Worth knowing what Google itself claims, and what it does not.
image-text-to-text · 10,64,703 downloads
Llama3-Med42-8B
Built by a hospital group, tuned on medical material, and open enough that you can look at what went into it instead of taking the claim on trust.
text-generation · 4,734 downloads
Lingshu-7B
Answers questions about a medical image rather than only about text. The kind of model behind an imaging demo at a conference stall.
image-text-to-text · 6,820 downloads
Llama3-OpenBioLLM-8B
Openly published, widely quoted in AI-in-medicine talks, and often shown with a benchmark score. Useful for seeing what such a score actually rests on.
text-generation · 2,666 downloads
HuatuoGPT-o1-8B
Trained to show its reasoning on a medical question, not just hand you an answer. Closer to how you would expect a registrar to present.
text-generation · 887 downloads
meditron-7b
Built by a university and unusually clear about exactly which medical texts it was trained on — which is what lets you judge where it will be weak.
text-generation · 1,426 downloads
BiomedNLP-BiomedBERT-base-uncased-abstract-fulltext
Trained only on PubMed. It does not chat; it sits underneath the search and screening tools you already use for a literature review.
fill-mask · 2,32,970 downloads
whisper-large-v3
Speech to text. Most doctors meet this one without knowing it, inside a dictation or transcription app. Worth knowing where your voice note goes.
automatic-speech-recognition · 49,54,242 downloads
Facts are read from each model’s published card on Hugging Face. Listing a model here is not an endorsement, a safety assessment, or a recommendation to use it in patient care.