NVIDIA releases Cosmos3-H-Dreams, a real-time generative surgical simulator trained on clinical cholecystectomy recordings, with four-step and two-step checkpoints

NVIDIA released Cosmos3-H-Dreams, a research model that generates surgical video in real time and responds to 28 action inputs. It is built on the 16B-parameter Cosmos3-H-Embodied-Nano and trained on video and kinematics from CMR Surgical Versius cholecystectomy (gallbladder removal) procedures. Two faster student checkpoints (four-step and two-step denoising) run interactively on a single RTX PRO 6000 GPU. NVIDIA also published a fine-tuning and distillation recipe so others can adapt it to their own datasets. It follows an earlier Cosmos-H-Dreams release.

The main audience is surgical robotics and simulation researchers, who get an open starting point and a reproducible recipe that runs on one workstation-class GPU. It could make it cheaper to prototype training environments for robotic surgery. Impact is narrow for now: this is a research release with no clinical deployment, and nothing here changes how surgery or surgical training is done today.

The "first real-time generative surgical simulator" claim is NVIDIA's own ("to our knowledge") and has not been independently checked. Realistic-looking video is not the same as accurate tissue behaviour, and NVIDIA itself admits visual realism does not prove biomechanical accuracy. Talk of autonomous surgical policies is speculation, and the post gives no validation by surgeons or comparison against physics-based simulators.

1. 01Independent evaluation by surgeons or academic groups, including whether the simulated tool and tissue responses match real biomechanics, not just how the video looks.
2. 02Hugging Face downloads and third-party fine-tunes on other procedures or datasets, which would show the recipe works outside NVIDIA's own setup.
3. 03Any published use of the model to train or test a surgical robot policy, and whether results transfer to real hardware or regulated clinical settings.