Apple

Protein structure prediction from general-purpose transformer blocks and flow matching, with no MSAs, pair representations, or triangle attention.

Apple's foundation model trained on behavioral signals from wearables, modeling 27 HealthKit metrics to improve predictions across 57 health tasks.

Wearable accelerometry foundation model distilled from a PPG encoder, predicting cardiovascular and health biomarkers from motion signals alone.

Motion foundation model for wearable accelerometry, trained with relative contrastive learning on 1B segments from 87,376 participants.

Self-supervised foundation model for wearable electrocardiograms, trained with participant-level contrastive learning on 141,000 participants.

Self-supervised foundation model for wearable photoplethysmography, trained with participant-level contrastive learning on 141,000 participants.