Google Research unveiled GlucoFM, a new foundation model designed to interpret continuous glucose monitoring (CGM) data with greater accuracy than existing approaches. The lightweight, self-supervised model uses a dual-stream architecture that separately processes slower glycemic trends and short-term glucose deviations, allowing it to capture the complex dynamics of blood sugar patterns over time. Built on 109,066 hours of unlabeled CGM data from multiple sources, GlucoFM can help predict metabolic conditions including diabetes risk, insulin resistance, and beta-cell dysfunction. In head-to-head evaluations against existing models like GluFormer and CGM-JEPA across four diverse cohorts, GlucoFM demonstrated significant performance advantages. The model achieved an average 5.8 percentage point improvement in precision-recall area under the curve (PR-AUC), a standard metric for evaluating prediction accuracy, and led in nearly all diabetes-risk and beta-cell-dysfunction prediction tasks. GlucoFM also excelled at predicting post-meal glucose responses with lower error rates than competitors, using data from Dexcom and Libre glucose monitors. A standout feature of GlucoFM is its transferability across datasets and its ability to adapt with minimal labeled data. The model's learned representations transfer effectively to new populations and cohorts, and the research demonstrated strong few-shot adaptation when working with extremely limited labeled data from new subjects. These capabilities suggest potential for deploying CGM-based metabolic prediction at scale, though clinical validation and regulatory pathways remain important next steps for real-world medical applications.