Microsoft cuts pathology-model compute 50x

Microsoft Research released GigaPath-Flash and GigaTIME-Flash, open-weight pathology models that cut whole-slide compute roughly 50× while keeping 97% of GigaPath’s predictive performance, and make spatial-proteomics inference roughly 6× faster with about eight times less memory, according to its research release. The models use 22-million-parameter tile encoders and ship under Apache 2.0, but Microsoft labels them research models rather than clinical tools; weights and code are available. For medical-AI teams, cheaper inference is a compute-budget signal, not a diagnostic green light—compare it against AMIE’s simulated clinical evidence gate before treating efficiency as deployment evidence.