NVIDIA Releases Cosmos3-H-Dreams for Surgical Video Generation

NVIDIA has published the nvidia/Cosmos3-H-Dreams model on HuggingFace with the image-to-video task. The model was published on 10/19/2026 and is designed to generate endoscopic video from a single camera frame and a stream of robot commands. It is a world model for CMR Surgical Versius surgical robotics, trained on clinical cholecystectomy data.

What it means

The model takes up 56.5 GB in the repository, with the largest weights file at 28.3 GB in safetensors format. The license is the NVIDIA Open Model License: commercial use is permitted, and you may create and distribute derivative models; when distributing, you must include the license text and a Notice file with attribution. Bypassing the model's technical limitations and safeguards is prohibited. The architecture is autoregressive and runs via nv-medtech. This is a narrowly specialized tool for simulating surgical operations, not a general-purpose video generator. In our video generators section, the closest image-to-video tools by task are LTX-Video and Maestro, but they solve different problems: LTX-Video generates general-purpose video and requires from 1 GB of VRAM, while Maestro is a studio for local video, images and music with 12—24 GB of VRAM. Cosmos3-H-Dreams is not intended for such scenarios. The developer does not specify VRAM requirements for running the model. For those working with video generation on their own hardware, this model is unlikely to become a working tool — its scope is limited to surgical simulation and robotics policy evaluation.