MedSigLIP

MedSigLIP is Google's open medical vision-language encoder, adapted from SigLIP-400M and fine-tuned on diverse medical imaging data including chest X-rays, pathology, dermatology, and ophthalmology images. It is the same encoder that powers image interpretation in the MedGemma generative models.

The architecture is a two-tower design: a 400M vision transformer plus a 400M text transformer, taking 448x448 images and short text prompts. Google published it with a model card on Hugging Face (google/medsiglip-448) and documented its creation date as July 9, 2025, alongside the MedGemma technical report (arXiv 2507.05201).

Honest limits: it is an encoder for classification, retrieval, and embedding - not a generative model. Benchmark claims (CXR14, dermatology AUC, EyePACS) are Google-reported. It is a research foundation, not a clinical device.