SleepFM-2 is a sleep‑foundation model built on 282,511 polysomnography (PSG) recordings, of which 235,865 were used for pre‑training, covering more than two million hours of multimodal physiology. Compared with the original SleepFM, SleepFM-2 improves disease prediction, sleep staging, detection of arousals, limb movements and respiratory events, and transfers to wearable sensing and subjective sleep phenotypes.\ \ When the PSG representation is combined with age, sex and BMI, the model meets a prespecified discrimination and significance criterion on two held‑out cohorts, correctly predicting 215 subsequently recorded electronic health record (EHR) phenotypes; for 155 phenotypes the PSG representation adds reproducible information beyond demographics. SleepFM-2 also outperforms a 480‑feature baseline derived from the same recordings.\ \ The disease scores reveal a reproducible principal component linked to reduced sigma‑band spatial coupling and increased hypnodensity entropy. The frozen encoder performs within the observed range of expert scorers for sleep events and transfers to wakeful EEG, headband, in‑ear EEG, wrist PPG and wrist accelerometry. It improves sleep staging across six accelerometry cohorts and achieves disease‑prediction performance in the UK Biobank comparable to models pretrained directly on accelerometry data.\ \ Finally, SleepFM-2 captures aspects of subjective sleep not recovered by conventional PSG summaries, particularly self‑reports of the recorded night. These findings demonstrate that multimodal sleep physiology can provide a transferable representation of human health across diseases, clinical tasks, sensors and subjective experience.\ \ Review