AI NewsSep 1, 2026Words 1504Read time≈ 4 min

Microsoft Releases GigaPath-Flash and GigaTIME-Flash for Efficient Pathology AI

Microsoft's open-weight pathology models cut whole-slide compute by 50× and spatial-proteomics memory use by nearly 8×.

AIOpen Source

Contents · 11

1. Two Smaller Models for Different Pathology Tasks
2. What the Benchmarks Show
3. The Compute Reduction at Population Scale
4. An Open-Weight Release With Practical Conditions
Frequently Asked Questions
What is the difference between GigaPath-Flash and GigaTIME-Flash?
How much smaller is GigaPath-Flash?
Are the models open source?
Can hospitals use the models for patient care?
Has the Flash research been peer reviewed?
Sources

Microsoft Research has released GigaPath-Flash and GigaTIME-Flash, two smaller pathology models designed to make whole-slide analysis and virtual spatial proteomics practical across much larger research cohorts.

The headline result is a sharp reduction in computation rather than a new clinical capability. GigaPath-Flash retained 97% of the original GigaPath’s average performance on two slide-classification benchmarks while using approximately 49.5 times fewer floating-point operations. GigaTIME-Flash exceeded the original GigaTIME’s average protein-prediction performance in the reported tests while running about six times faster and using nearly eight times less GPU memory at the largest tested batch size.

Microsoft published its announcement on August 31, 2026. The accompanying technical report was first submitted to arXiv on July 20 and revised on July 22. The models are joint work from Microsoft Research, the University of Washington, Providence Genomics, Providence Cancer Institute, and Providence Research Network.

1. Two Smaller Models for Different Pathology Tasks

GigaPath-Flash converts a gigapixel pathology slide into representations that can support downstream research tasks such as tumor classification. Like the original GigaPath, it processes a slide in two stages: a tile encoder extracts local visual features, and a slide encoder combines those features with their spatial positions to model tissue-wide context.

The Flash version replaces GigaPath’s approximately one-billion-parameter ViT-g tile encoder with a 22-million-parameter ViT-S/16 model. Each 224-by-224-pixel tile becomes a 384-dimensional embedding. A separate 12-layer LongNet slide encoder, containing approximately 21 million parameters, then contextualizes the embeddings using dilated attention that scales linearly with the number of tiles.

The smaller encoder was distilled from the frozen billion-parameter model rather than trained independently. The researchers used a DINOv2 self-supervised objective on whole-slide images from Providence, transferring representations from the larger teacher into the compact student. The slide encoder was subsequently pretrained with masked autoencoding, learning to reconstruct a tile’s representation from the surrounding slide.

GigaTIME-Flash serves a different purpose. It accepts routine hematoxylin-and-eosin, or H&E, image tiles and predicts spatial maps for 21 protein channels normally measured through multiplex immunofluorescence. Its released implementation produces 23 output channels, including two background channels.

The original GigaTIME used a nine-million-parameter convolutional U-Net++ architecture. GigaTIME-Flash instead combines the distilled ViT-S encoder with a two-million-parameter convolutional decoder, for a total of about 23.8 million parameters. Features from four transformer layers feed successive decoder stages, preserving spatial detail while most computation occurs in the lower-resolution token representation.

Only LoRA adapters and the decoder are trained during adaptation; the other encoder parameters remain frozen. The reported configuration uses LoRA rank 8, scaling factor 16, and dropout of 0.1. Training ran for 300 epochs on NVIDIA A100 hardware using the same data as the original GigaTIME comparison.

2. What the Benchmarks Show

GigaPath-Flash was evaluated on two public slide-level classification datasets. PANDA tests six-class prostate-cancer grade prediction, while EBRAINS covers 30 brain-tumor subtypes.

For an EBRAINS slide containing 31,469 tiles, the researchers estimated GigaPath-Flash at 290.3 TFLOPs, compared with 14,367.3 TFLOPs for the original GigaPath. Its PANDA quadratic-weighted kappa was 0.947, versus 0.965 for GigaPath, while EBRAINS balanced accuracy was 0.705, versus 0.741.

Averaging those two scores produced 0.826 for GigaPath-Flash and 0.853 for GigaPath. That is the basis for Microsoft’s claim that the smaller model preserves approximately 97% of GigaPath’s average performance.

GigaPath-Flash also exceeded all tile-only baselines included in the study on both benchmarks. Those models required an additional attention-based multiple-instance-learning component to combine independent tile representations. Among the evaluated models pretrained at whole-slide level, GigaPath-Flash had the lowest estimated inference cost.

These comparisons remain narrower than a general pathology evaluation. The researchers created custom train, validation, and test splits, trained each downstream model for five epochs, and reported one run per dataset. The results therefore support a controlled efficiency comparison, not a claim that GigaPath-Flash is universally superior across pathology tasks.

GigaTIME-Flash was tested on 9,204 registered tiles from five lung-adenocarcinoma patients and on out-of-distribution tissue microarrays covering brain, breast, colon, and lung cancers. Each external cancer cohort contained samples from approximately 10 to 20 patients.

Performance was measured using Pearson correlation between predicted and experimentally measured protein activation, aggregated over windows corresponding approximately to individual cells. GigaTIME-Flash achieved a higher mean correlation than GigaTIME on the original test set and all four external cohorts.

The gains were not uniform across proteins. GigaTIME-Flash improved predictions for several nuclear, epithelial, myeloid, proliferation, and apoptosis markers. The original convolutional model remained competitive for some vascular, stromal, and sparse lymphoid patterns, including CD34, Transgelin, Actin, and CD20.

3. The Compute Reduction at Population Scale

For a 256-by-256-pixel input, GigaTIME-Flash requires 14.9 GFLOPs, compared with 69.1 GFLOPs for GigaTIME—a reduction of approximately 4.6 times.

Its advantage grew with batching on an NVIDIA A100. Throughput rose from 60.3 tiles per second at batch size 1 to 1,679.2 tiles per second at batch size 128. The original GigaTIME plateaued at roughly 390 tiles per second.

At batch size 128, peak GPU memory was 2.16 GB for GigaTIME-Flash and 16.68 GB for GigaTIME. The newer model is larger by parameter count, but its transformer backbone makes greater use of matrix operations that parallelize efficiently on modern GPUs, while the original model performs dense convolutions and nested skip operations on high-resolution feature maps.

Microsoft estimates that generating virtual multiplex-immunofluorescence maps for one million slides would require approximately 70 days on one A100 with GigaTIME-Flash, assuming 10,000 tiles per slide and batch size 128. Its corresponding estimate for GigaTIME is about 300 days. For 100,000 slides, the estimates are seven days and 30 days, respectively.

Those figures are projections rather than measured end-to-end cohort runs. Actual times depend on slide dimensions, tiling resolution, preprocessing, storage throughput, implementation, and hardware.

The scale matters because the original GigaTIME study applied virtual protein prediction to 14,256 Providence patients across 24 cancer types and 306 subtypes. It generated 299,376 virtual protein maps and reported 1,234 statistically significant protein–biomarker associations. Lowering inference cost makes it more feasible to repeat such analyses across alternative biomarkers, patient subsets, and validation cohorts.

4. An Open-Weight Release With Practical Conditions

Microsoft and its collaborators released both Flash checkpoints and their code under the Apache License 2.0. The GigaPath-Flash weights are hosted through the Prov-GigaPath organization on Hugging Face, while GigaTIME-Flash is distributed through the Prov-GigaTIME organization. The public GitHub repositories include loading instructions and inference notebooks.

The weights are gated: users must sign in, agree to the repository terms, and provide a read-only Hugging Face token before downloading them. This means the artifacts are openly licensed and inspectable after access is granted, but they are not anonymous, one-click downloads.

There is also a distinction between the license and the stated intended use. Apache 2.0 is a permissive software license, but both model cards describe the checkpoints as research resources and say deployed uses, commercial or otherwise, are outside their intended scope. The materials do not explain how that model-card guidance should be reconciled with every use permitted by the license, so organizations contemplating deployment should obtain clarification rather than treating “Apache 2.0” as evidence of clinical or operational approval.

Neither model is validated for diagnosis, prognosis, treatment selection, or other patient-care decisions. The authors call for multi-institutional and prospective validation across scanners, institutions, patient populations, and downstream tasks.

That caution is supported by broader pathology research. An independent 2026 study found that pathology foundation models can encode hospital-specific staining, preparation, and scanner signatures, allowing downstream systems to rely on institutional artifacts instead of biological features. That study evaluated the original Prov-GigaPath among a wider set of models, not GigaPath-Flash, so it does not establish a defect in the new release. It does show why lower compute and favorable benchmark scores cannot substitute for cross-site validation.

Frequently Asked Questions

What is the difference between GigaPath-Flash and GigaTIME-Flash?

GigaPath-Flash creates contextualized representations of whole pathology slides. GigaTIME-Flash predicts spatial protein-expression maps from routine H&E images.

How much smaller is GigaPath-Flash?

Its tile encoder has approximately 22 million parameters, compared with about one billion in the original GigaPath. Its slide encoder has approximately 21 million parameters, compared with 86 million previously.

Are the models open source?

The code and weights are published under Apache 2.0, but checkpoint downloads require a Hugging Face account and acceptance of access terms. “Open-weight” is therefore the most precise description.

Can hospitals use the models for patient care?

No. Microsoft and the model cards state that the models are research releases and are not intended or validated for clinical diagnosis, prognosis, treatment selection, or other care decisions.

Has the Flash research been peer reviewed?

The Flash results are currently documented in an arXiv preprint. The original GigaPath and GigaTIME research appeared in Nature and Cell, respectively.