Google Research has unveiled GlucoFM, a self-supervised foundation model designed to analyze continuous glucose monitoring (CGM) data and predict metabolic health conditions including diabetes risk, insulin resistance, and beta-cell dysfunction. The model employs a dual-stream architecture that separates slower glycemic trends from short-term deviations, allowing it to capture both baseline glucose patterns and transient fluctuations caused by meals, activity, or sensor variations. Trained on over 109,000 hours of unlabeled CGM data from multiple cohorts, GlucoFM significantly outperformed existing models in clinical evaluations. Across 14 cohort-task evaluations spanning seven prediction tasks, GlucoFM achieved PR-AUC scores 5.8 percentage points higher on average than the best-performing GluFormer variant, while also demonstrating superior performance in postprandial glucose response forecasting across multiple CGM devices. Beyond raw performance metrics, GlucoFM shows strong cross-dataset transfer capabilities and effective few-shot adaptation when labeled data are limited—a critical advantage in healthcare applications where quality clinical labels are expensive and sparse. The model's ability to learn from diverse, unlabeled glucose data could accelerate early detection of metabolic disorders and enable personalized health interventions.