Microsoft Research and Paige unveil PRISM2, a language-aware pathology foundation model

Microsoft Research, working with Paige (now part of Tempus), has released PRISM2, a pathology foundation model that pairs tissue-image analysis with the language of pathology reports themselves — a departure from prior models that learned only from images.

What's new

Microsoft published the work on August 4, 2026, on its Source blog. PRISM2 combines whole-slide tissue images with the diagnostic language pathologists use to describe them, training the model to connect visual patterns directly to the terminology and reasoning found in real pathology reports. Microsoft frames the design choice plainly: "Pathology is not only a visual discipline. It is also a language-driven one."

According to Microsoft, PRISM2 matched or exceeded the performance of specialized, task-specific cancer-detection systems across multiple benchmarks — without requiring a separate purpose-built model for each diagnostic task, which has been the norm in computational pathology. The model weights have been made publicly available on Hugging Face for research use.

Context

Computational pathology has largely followed a narrow-model paradigm: a distinct system trained for each specific detection task, such as identifying a particular cancer subtype in a specific tissue type. That approach requires extensive labeled data and separate engineering effort per task. Earlier pathology foundation models improved on this by learning general visual representations from large image datasets, but still treated the accompanying diagnostic text as, at best, a secondary signal. PRISM2 builds directly on this foundation-model lineage but elevates report language to a first-class training signal alongside the images.

Why it matters

A single model that generalizes across diagnostic tasks by learning from both images and the language pathologists already use has practical implications for how quickly pathology AI can be adapted to new cancer types or diagnostic questions, since it reduces the need to build and validate a bespoke model per task. Publishing the weights on Hugging Face also lowers the barrier for hospitals, research labs, and other developers to build on or evaluate PRISM2 directly rather than relying solely on Microsoft's own deployment, which could accelerate independent validation of the model's claimed generalization across benchmarks.