AI News Feed
Market watch
LG UltraFine Evo 32U990A-S Hits Record-Low $1,099 on Amazon, TechRadar ReportsNewegg Labor Day Deal: 32GB Kingston Fury Beast DDR4-3200 Kit Nearly Half Price at $179.99 with CodeGalaxy Z Fold 8 inner screen corners reported with excessive giveHuawei unveils Mate XT 2 tri-fold with privacy-screen edition as Apple prepares folding iPhoneWall Street Analysts Recommend Nvidia, Uber and Marvell for Long-Term GrowthLenovo, AMD and Huawei Put Local Large-Model AI at Center of New Devices
LG UltraFine Evo 32U990A-S Hits Record-Low $1,099 on Amazon, TechRadar ReportsNewegg Labor Day Deal: 32GB Kingston Fury Beast DDR4-3200 Kit Nearly Half Price at $179.99 with CodeGalaxy Z Fold 8 inner screen corners reported with excessive giveHuawei unveils Mate XT 2 tri-fold with privacy-screen edition as Apple prepares folding iPhoneWall Street Analysts Recommend Nvidia, Uber and Marvell for Long-Term GrowthLenovo, AMD and Huawei Put Local Large-Model AI at Center of New Devices
Large Language Models

Google Research and UNSW Sydney Release GlucoFM, a 0.72M-Parameter Foundation Model for Continuous Glucose Monitoring

Google Research and UNSW Sydney have released GlucoFM, a 0.72M-parameter dual-stream foundation model for continuous glucose monitoring that outperforms existing CGM baselines on clinical prediction tasks.

AI News Feed · 2026-08-27 21:28

Existing CGM foundation models such as CGMformer, GluFormer and CGM-JEPA encode a glucose trace as a single entangled sequence. GlucoFM, by contrast, decomposes it into two streams while keeping the observation mask intact, and pretrains with two JEPA-style latent objectives.

The model was pretrained on 109,066 hours of unlabeled CGM data from 477 subjects, using a single NVIDIA H100. It has 0.72 million trainable parameters and 1.18 million total parameters, with a three-layer Transformer encoder. The pretraining ran for 120 epochs.

Under subject-disjoint linear probing across four cohorts and seven tasks, GlucoFM achieved 58.8 task-averaged PR-AUC versus 54.7 for the strongest CGM-specific baseline, a relative improvement of about 7.5%. It led PR-AUC on all diabetes-risk and beta-cell-dysfunction evaluations and on three of four insulin-resistance evaluations, and ranked first on 21 of 24 cross-dataset transfer evaluations. For two-hour postprandial glycemic response forecasting, it reached 21.88 mg/dL MAE with full context, compared to 22.90 for the best baseline and 27.69 for a train-fold mean.

The 0.72M-parameter model also outperformed a 135M-parameter GluFormer and a 385M-parameter MOMENT on task-averaged PR-AUC. Trained on only 20% of the corpus, it already matched CGM baselines trained on the full corpus. It beat a seven-day GMI threshold rule on macro-F1 by 7.4 points on Stanford and 17.4 points on CGMacros-Dexcom.

The architecture aligns recordings to a fixed 24-hour grid at 5-minute intervals, preserving the absolute circadian start index. A causal, mask-aware learnable Gaussian filter splits the signal into filtered trend (state stream) and masked residual (event stream). Both streams are tokenized into 24 one-hour patches and fused into 128-dimensional tokens with circular time-of-day features. Pretraining uses two JEPA-style objectives: masked contextual latent prediction over 50–60% of patches against an EMA teacher, and next-patch state/event dynamics prediction.

GlucoFM is research infrastructure, not a clinical or consumer product. The research team states that it has not been cleared or approved by any regulatory authority and is not intended to diagnose, treat, cure or prevent disease. All evaluations are retrospective, the largest pretraining cohort is non-public, and no checkpoint has shipped as of August 26, 2026. The paper commits to releasing code and reproducibility scripts. The trained recipe is reproducible on a single NVIDIA H100, and 24-hour-window inference can run on a CPU container or on-device.