Google Research and UNSW Sydney just dropped a foundation model for continuous glucose monitoring that’s small enough to run on a phone, yet outperforms models hundreds of times its size. The trick isn’t more compute — it’s recognizing that a glucose trace contains two fundamentally different signals tangled together.
GlucoFM separates each 24-hour glucose recording into a slow physiological “state” stream and a transient “event” stream, then trains a 0.72M-parameter Transformer on both using JEPA-style self-supervised objectives. After pretraining on 109,066 hours of unlabeled CGM data from 477 subjects across five cohorts, it hit 58.8 task-averaged PR-AUC across 14 cohort–task evaluations. The strongest CGM-specific baseline retrained on the same corpus managed 54.7. That’s a 7.5% relative gain — from a model roughly 187 times smaller than the 135M-parameter GluFormer and 535 times smaller than the 385M MOMENT it also beat.
The paper notes a curious detail: dense interpolation of missing glucose readings actually hurts performance compared to simply preserving the observation mask. The model learns better when it knows what it doesn’t know. Pretraining ran on a single H100 for 120 epochs, and inference on a 24-hour window fits in a CPU container or on-device. Any research group with a CGM dataset can reproduce the recipe.
Before anyone gets excited about shipping this to patients: the researchers are explicit that GlucoFM is a research prototype with no regulatory clearance and no public checkpoint as of late August 2026. Every evaluation is retrospective, and the largest pretraining cohort isn’t public. Still, the strongest gains landed exactly where they matter clinically — diabetes risk, beta-cell dysfunction, and insulin resistance. Trained on just 20% of the corpus, GlucoFM already matched baselines that used the full dataset. That efficiency signal, more than any single benchmark number, suggests the state-event split is doing real work rather than just surfing on scale.
