PRISM2 gives pathology AI a broader research target: learning from both microscope images and the language pathologists use in real reports, instead of building a separate model for every cancer-detection task.
Microsoft Research and Paige, now part of Tempus, described the system in Nature Medicine. The team trained the pathology foundation model on tissue images paired with pathology-report text so it could answer prompts and support different downstream research uses.
In tests cited by Microsoft, PRISM2 matched or exceeded specialized systems on several benchmark tasks, including prostate cancer, breast cancer and breast lymph node metastasis detection. The point is not that the model is ready to diagnose patients by itself, but that a single promptable model may reduce the amount of rebuilding needed for each new pathology application.
Microsoft says the full model weights are available on Hugging Face for research use. That makes the near-term impact most relevant to labs and tool builders studying how multimodal medical AI can support clinicians while still requiring validation before clinical deployment.