PRISM2 trains on 2.3M pathology slides

Microsoft Research and Paige released PRISM2, a 4.4-billion-parameter pathology foundation model trained on 2.3 million whole-slide images, 685,507 specimens, and 14 million question-answer pairs, according to the research-only model card and weights. The accompanying Nature Medicine paper reports that prompt-based inference matched or exceeded three specialized cancer-detection products without task-specific training. Unlike the commercial machinery described in OpenAI’s life-sciences platform strategy, these gated weights prohibit commercial and clinical use, so builders should file PRISM2 as a research baseline—not a deployable diagnostic shortcut.