Google Research and the University of New South Wales in Sydney have introduced GlucoFM, a self-supervised foundation model built for continuous glucose monitoring. The model processes glucose signals through a dual-stream setup that separates slower physiological "state" signals from transient "event" signals. According to the release, GlucoFM has just 720,000 parameters and was pretrained on a single NVIDIA H100 GPU, pointing to a relatively lightweight training footprint compared with larger deep learning systems. In evaluations spanning 14 cohort-task settings, the model posted an average PR-AUC of 58.8 and outperformed baseline models. The research team said GlucoFM remains a research prototype and has not been cleared by any regulator. It is not intended for disease diagnosis or treatment. The team also said it plans to release the code and reproduction scripts, allowing groups with glucose monitoring datasets to run inference on CPUs or on-device systems using the same approach. Techub cited MarkTechPost as the source for the update.