NVIDIA introduced Cosmos3-H-Dreams on October 20, 2026: generative surgical simulation trained on video and kinematics from procedures using Versius. The company reports interactive execution on one RTX PRO 6000. This is a research system without clinical validation for patient care; SIExpo has not independently tested its quality or speed.
From an initial scene to controlled video
The public implementation takes an initial image and instrument actions, generating video in chunks. It can replay a recorded trajectory or accept VR-controller input through a browser. It requires Python 3.12, a compatible GPU and separately supplied weights and user assets. Example assets have separate provenance and licensing records; the code license does not replace those conditions.
Our technical explanation is that controlling the next generated image should be distinguished from playing a prerecorded video. Given the same starting point, a different action should lead to the corresponding change. For an assessment, we would retain input commands and generated frames so that a reviewer could compare the cause with the visible result. This is our proposed protocol, which we have not executed against the model or used to establish its reliability.
What the execution layer provides
FlashDreams is the separate inference and serving library for interactive autoregressive video and world models. Its repository specifies Apache-2.0 with separate terms for third-party components. That explains the role of the software layer; it does not provide independent medical confirmation of the surgical simulator’s behavior.
In such an environment, we would examine more than a convincing frame. For example, we would check whether a movement produces the expected instrument direction and whether the scene preserves its structure over a sequence of commands. Expert review and measurements need to remain separate from the impression of realism. This illustrative example does not claim that we observed a defect or a successful clinical procedure, and it should not be read as clinical advice.
Why the boundary with medical practice matters
Our assessment is that the value of a platform of this kind needs to be demonstrated for a defined research task. A useful demonstration can help researchers formulate questions without establishing answers about patient safety. We do not infer authorization for autonomous operations, improved treatment outcomes or universal accuracy of simulated behavior from the published video.
For a future pilot, we would request explicit criteria: what gets evaluated, who checks it and which cases lie outside the scope. We would also separate development data from evaluation data and retain failed continuations. This would make the research assessable on its own terms and show which conclusions are supported by actual measurements. Publishing tools can make inspection more accessible, but the basis for a particular use still needs to come from the results of that inspection. A compelling visual result is a starting point for that work rather than its completion.
Editorial verification: 2026-10-21T06:28:23.703396+00:00 (UTC). Cover: conceptual AI illustration, not a documentary photograph.
