Google Splits the Glucose Curve into Two Streams — GlucoFM Reads Out Metabolic Health

One Glucose Curve, Split Into Two Streams

On August 26, 2026, Google Research and UNSW Sydney jointly released GlucoFM, a self-supervised foundation model (pretrained on large-scale data and adaptable to multiple tasks). It splits continuous glucose monitoring data into two independent streams—a slow physiological trend and short-term fluctuations—modeled separately.

The core design of GlucoFM is a dual-stream encoder. One stream handles a low-frequency state component, representing the slow baseline changes in glucose; the other handles a residual event component, capturing short-term deviations from meals, exercise, or sensor artifacts. Inputs are aligned to a 24-hour, 5-minute grid with an observation mask preserved so the model knows which time points have real data. This structure differs from existing CGM models: CGMformer, GluFormer, and CGM-JEPA all treat glucose sequences as a single stream, without explicitly separating fast and slow components.

Evaluation results come from 4 cohorts, 7 clinical tasks, and 14 cohort-task combinations in total. GlucoFM's average PR-AUC is 4.1 percentage points higher than the strongest CGM-specific baseline retrained on the same pretraining corpus (54.7 → 58.8, approximately +7.5% relative). It leads across the board on diabetes risk assessment and β-cell dysfunction; for insulin resistance, it leads on 3 of 4 evaluations. Testing also covers hyperlipidemia, hypoglycemia, obesity, and glucotype classification. The project is currently in the research stage, and weights have not been released.

CGM Data Isn't Uniform—It Learns by Splitting Into Two Streams

GlucoFM's core idea is simple: the slow trends and short fluctuations in a glucose curve shouldn't be processed through the same channel.

A continuous glucose monitor records interstitial glucose concentration every few minutes—CGM (continuous glucose monitoring, using a subcutaneous sensor to record glucose values every few minutes). But this curve isn't uniform—it contains a slowly changing baseline (such as fasting and overnight patterns) and rapidly fluctuating deviations (from meals, exercise, or sensor noise). Previous mainstream models used a single-stream encoder to process the entire sequence. Researchers at Google Research and UNSW Sydney felt this was wrong: baseline patterns and event patterns are two different physiological signals, and learning them together creates interference.

GlucoFM decomposes a CGM trajectory into two streams: one captures the slow physiological "state," the other captures transient "events." Both streams retain timestamps and missing-value markers, then undergo self-supervised pretraining with two separate JEPA-style objective functions—self-supervised learning (a training approach that lets the model learn from the data's own structure, without human-labeled tags). This way, the model learns during pretraining to distinguish "what is this person's fasting glucose baseline" from "how much fluctuation did that meal just cause."

The direct benefit of the dual-stream design is reduced demand for labeled data. Traditional methods for predicting diabetes risk or insulin resistance typically require large amounts of clinically labeled CGM records. GlucoFM first pretrains on the unlabeled data listed above, then fine-tunes (further training a pretrained model on a small amount of task-specific data to adapt it to a new task) with just a few labeled samples to transfer to new tasks. Google reports in the paper that GlucoFM achieves performance gains across 7 clinical prediction tasks, including diabetes risk assessment, insulin resistance, and β-cell dysfunction.

The key point: GlucoFM didn't win by using a larger model—according to industry media Agent5, its encoder has only 0.72M parameters. The official blog did not disclose the parameter count, nor has it released full weights, and there are no third-party replication results. Whether the dual-stream strategy outperforms single-stream in all scenarios still needs independent verification.

0.72M Parameters Beat a Stronger Baseline. How?

Google Research's GlucoFM has only 0.72M trainable parameters, according to Agent5. That is almost nothing for a foundation model. Yet across 14 cohort-task evaluations, its average PR-AUC (area under the precision-recall curve, measuring the quality of prediction ranking) beats the strongest CGM-specific baseline by 4.1 percentage points.

Previous foundation models (pretrained on large-scale data, adaptable to multiple tasks) for continuous glucose monitoring (CGM) data learned from the entire glucose curve as a single stream. GlucoFM goes the other way: self-supervised (no manual labels needed—finds patterns directly from the data) pretraining splits the curve into a slow physiological "state" and brief "event" fluctuations. Each stream only needs to learn a simpler task.

Across four cohorts and seven clinical tasks, GlucoFM ranks first in all evaluations for diabetes risk and β-cell dysfunction, and wins 3 of 4 insulin resistance evaluations. The pattern holds across devices. Separate models built for Dexcom and Libre, the two leading CGM devices, hit the lowest mean absolute error for postprandial glucose prediction at 21.88 mg/dL, below the best baseline's 22.90. That result comes from Google's own report and has not yet been independently replicated.

Cross-Cohort Transfer: Can Minimal Labels Sustain Metabolic Prediction?

Can GlucoFM hold up when it moves from pretraining to a new cohort where labels are scarce? The judgment rests on a report card spanning four cohorts and seven tasks.

Just looking at a glucose curve won't tell you how severe someone's insulin resistance is. Judging such metabolic indicators requires calibration through clinical examinations; the paper states plainly that high-quality clinical labels are scarce and expensive. In practice, collecting labeled CGM data costs far more than collecting the glucose curves themselves.

The approach: learn the patterns cleanly first.

GlucoFM is a foundation model (pretrained on massive data, adaptable to multiple downstream tasks) that uses self-supervised (no manual labels—finds patterns directly from the data itself) pretraining, predicting daily context and temporal evolution of glucose curves without looking at labels. After pretraining, it adapts to a new cohort with just a small number of labeled samples—that's what few-shot (learning a new task with very few labels) means.

The report card spans four cohorts and seven clinical tasks (diabetes risk, insulin resistance, β-cell dysfunction, hyperlipidemia, hypoglycemia, obesity, and glucotype), for 14 cohort-task combinations in total. Against the strongest CGM-specific baseline retrained on the same corpus, GlucoFM's average PR-AUC (a metric for evaluating prediction ranking quality, more reliable with imbalanced samples) is 4.1 percentage points higher; it leads every evaluation for diabetes risk and β-cell dysfunction, and wins 3 of 4 insulin resistance evaluations. According to Agent5, it achieves this with just 0.72M parameters (the total number of tunable weights in a model; the smaller, the more compute-efficient). Caveat: the numbers come from Google's own evaluation; as of publication, no third-party replication has appeared.

What to watch: everything going forward comes down to that word "transfer." GlucoFM is still in the research stage, and model weights haven't been released. Once weights are actually out, watch whether third parties can reproduce the advantage on their own CGM cohorts when labeled samples are cut to a minimum—and whether the dual-stream split gets adopted by subsequent CGM models. If the few-shot advantage disappears on new cohorts, or if later models revert to single-stream, this direction will need revision.

Don't Look for a Demo Yet—Weights Haven't Been Released

On the day GlucoFM was released, the paper and blog were both public, but the model weights weren't uploaded. What's possible right now: read the paper, watch the repo, and understand one key distinction.

According to Agent5 and explainx, Google Research and UNSW Sydney released GlucoFM on August 26, but the weights (trained parameter files; without them the model can't run) are marked TBD—to be released. The official blog's Quick links only include a Paper link, with no download entry. What you can do first is read the paper: open the Paper, look at how the dual-stream structure splits the CGM (continuous glucose monitoring; a subcutaneous sensor measures glucose every few minutes) curve into a slow "state stream" and a fast "event stream," while preserving temporal information and missing values (time points where no measurement was taken).

Read the paper with comparison in mind: CGMformer, GluFormer, and CGM-JEPA all treat the entire glucose curve as a single sequence; GlucoFM splits it into two. You can understand this difference without running any code. If you wear a CGM device, pull up your own week-long curve and separate the smooth baseline portions from the post-meal or activity-driven spikes—that's the intuition behind dual streams.

Don't rush to memorize the numbers. The official claim is that PR-AUC (a prediction ranking quality metric; higher is better) averages 4.1 percentage points higher than the best CGM-specific baseline, but that's against their own trained baseline—no third-party replication yet. Once weights or code are public, see if the number can be reproduced. Until weights are released, the most valuable action is just one: remember the dual-stream approach and the fact of the lightweight design.