Microsoft Research has released two pathology foundation models under the Apache 2.0 licence, with weights and code on Hugging Face. The number worth stopping on is the size. GigaPath-Flash pairs a 22 million parameter tile encoder with a 21 million parameter slide encoder, and the tile encoder was distilled from a billion-parameter teacher.
That is the distillation claim stated in a field where compute is not a convenience. It decides which questions get asked at all.
Pathology is a compute problem before it is a modelling problem#
A whole-slide image often exceeds a gigapixel, and running a foundation model across one slide means processing thousands of tiles. Hospitals produce millions of slides a year.
Microsoft's description of the bottleneck is the useful part, because it is not the one people usually give. Population-scale discovery, it says, "is not a single model run": it is repeated cycles of feature extraction, statistical analysis, hypothesis testing and validation across patient subgroups, biomarkers and clinical endpoints.
So the cost does not scale with the number of slides. It scales with the number of slides multiplied by the number of questions you want to ask of them, and the second factor is the one a researcher gives up first. That is the same economics as which questions researchers can afford to ask, arriving here as a concrete instance rather than an argument.
What the distillation actually bought, and what it cost#
On slide-level classification, measured on PANDA prostate grading and EBRAINS brain tumour subtyping, Microsoft reports GigaPath-Flash scoring within 3 percent of the original GigaPath at roughly 50 times less compute, and holding the lowest inference cost among whole-slide pretrained models.
Read that as a trade rather than a free win, because it is one. Three percent of predictive performance went away. What came back was a model an order of magnitude smaller than its teacher, and the compute figure is not a rounding difference: 50 times is the gap between an experiment you run once and an experiment you run across a cohort, repeatedly, while changing your mind about the question.
The slide encoder is a LongNet that contextualises every tile embedding through dilated attention, which Microsoft says scales linearly with tile count. That linearity is doing as much work in this result as the distillation is, because a quadratic term would put the saving back at the top end where it is needed most.
The second model went the other way#
The two halves of this release did not come out the same, and the summary flattens them.
GigaTIME-Flash swaps the original GigaTIME's CNN backbone for the GigaPath-Flash encoder, adds a lightweight convolutional decoder, and is fine-tuned with LoRA adapters that keep the pretrained encoder weights largely frozen. Microsoft reports it running about 6 times faster with about 8 times less memory, and matching or improving on the original across in-distribution and out-of-distribution cohorts spanning brain, breast, colon and lung cancers.
So one model lost 3 percent and the other gained, from the same backbone swap. The gains were most pronounced on out-of-distribution data, and Microsoft's reading of that is offered as a hypothesis rather than a result: it says the foundation model backbone appears to improve generalisation to tissue types the model has not seen. Worth watching, not yet worth repeating as fact.
The throughput table is an estimate and says so#
Microsoft publishes wall-clock projections for generating virtual spatial proteomics across cohorts of different sizes, and reports GigaTIME-Flash scaling past 1,600 tiles per second.
The footing under those numbers is stated plainly: they assume roughly 10,000 tiles per slide, a batch size of 128 and a single NVIDIA A100, and actual runtime depends on slide size, tiling resolution and hardware. That is an estimate on named assumptions, which is the good version of this disclosure. Treat the shape of the curve as the claim and your own hardware as the number.
What it is not#
Both are research models. Microsoft states before anything else that they are not intended or validated for clinical use, naming diagnosis, prognosis, treatment selection and other patient-care decisions, and warns that performance may vary across datasets, scanners, institutions and populations. It calls this an early research release, says its evaluations cover a limited set of benchmarks and cohorts, and says downstream clinical applications will need additional multi-institutional and prospective validation.
That is the disclosure doing its job, and it draws the line cleanly. This is a tool for research groups asking questions of tissue at population scale, not something approaching a clinic. If you want the wider picture of where that kind of work is actually paying off, AI on hard scientific problems is the longer view.
The part that travels beyond pathology is the release itself. A billion-parameter model compressed to 22 million, given away under Apache 2.0 with the limitations printed above the results, is a better argument for open weights than most papers written about them.
Discussion