NVIDIA's Cosmos-H-Dreams Distills Surgical World Model to 160 FPS on One GPU
NVIDIA's Cosmos-H-Dreams Distills Surgical World Model to 160 FPS on One GPU
AI video model releases
Sign in to create alerts.
NVIDIA released Cosmos-H-Dreams, a real-time surgical world model distilled from the offline Cosmos-H-Surgical-Simulator, running at about 160 fps on a single RTX PRO 6000 GPU versus roughly 10 fps for its teacher model.
The model has no physics engine: it takes one initial frame plus a stream of robot arm commands and predicts the next block of 12 video frames, letting a keyboard, a Meta Quest headset, or a learned policy steer the generated scene live.
Speed comes from self-forcing distillation (training on its own generated history) combined with FlashDreams, NVIDIA's streaming inference library that caches repeated GPU computation, producing each frame in 2 to 4 denoising steps at 288x512 resolution and 10 Hz control rate.
The released checkpoint only handles tabletop suturing on the da Vinci Research Kit; NVIDIA and CMR Surgical wired it into the Versius surgeon controller and demoed it live at a robotic surgery conference in Florida in late July 2026, explicitly for research only and not for clinical use.
NVIDIA has not yet published accuracy results (tool-tip pose, drift over long rollouts, agreement with real policy outcomes), leaving open whether the fast distilled model is trustworthy enough to actually evaluate surgical policies.
cho.sh