Microsoft Debuts Efficient GigaPath-Flash AI Models

Microsoft Research, in collaboration with the University of Washington and Providence, has released GigaPath-Flash and GigaTIME-Flash. These open-weight models, distributed under the Apache 2.0 license on HuggingFace, are highly efficient versions of the company's previous pathology foundation models. By distilling the original billion-parameter GigaPath ViT-g teacher model, the researchers created a compact 22-million-parameter ViT-S tile encoder. This backbone is paired with a 21-million-parameter LongNet slide encoder to form GigaPath-Flash, which scales linearly to handle whole-slide images.

The efficiency gains do not come at the expense of accuracy. On slide-level classification benchmarks, including PANDA prostate grading and EBRAINS brain tumor subtyping, GigaPath-Flash achieved competitive results. It scored within 3 percent of the original GigaPath model while requiring roughly 50 times less computational power, delivering the lowest inference cost of any whole-slide pretrained model.

Meanwhile, GigaTIME-Flash adapts this technology for spatial proteomics. It replaces the convolutional neural network backbone of the original GigaTIME with the new ViT-S encoder and a lightweight convolutional decoder, fine-tuned using LoRA adapters. This configuration runs approximately six times faster and uses about eight times less memory than its predecessor, while matching or improving predictive performance across brain, breast, colon, and lung cancer datasets. The model can scale to process more than 1,600 tiles per second.

For medical researchers, these optimizations make population-scale studies feasible on standard hardware. Processing a cohort of one million slides, assuming 10,000 tiles per slide on a single Nvidia A100 GPU, would take about 300 GPU-days using the original GigaTIME. With GigaTIME-Flash, that timeline shrinks to just 70 GPU-days. For smaller cohorts of 1,000 slides, processing takes only two GPU-hours instead of seven. While these models are strictly for research and not validated for clinical diagnostic use, they allow scientists to run repeated, affordable analyses across massive patient cohorts to discover novel disease biomarkers.