Microsoft Research has released GigaPath-Flash and GigaTIME-Flash, efficient variants of its pathology foundation models designed to enable population-scale discovery by drastically reducing computational requirements.
- GigaPath-Flash uses a distilled 22M-parameter ViT-S tile encoder and a 21M-parameter LongNet slide encoder, achieving competitive performance within 3% of the original GigaPath at roughly 50 times less compute.
- GigaTIME-Flash replaces the original CNN backbone with the GigaPath-Flash ViT-S encoder to predict spatial proteomics from H&E images, matching or improving upon the original model's quality on out-of-distribution cohorts.
- The Flash family models are released under the Apache 2.0 license and support repeated analyses across larger patient cohorts for investigating disease biology and biomarkers.
These open models allow researchers to analyze larger cohorts and run more experiments affordably, moving computational pathology toward practical population-scale discovery.