SSyncDev

Cosmos3-H-Embodied-Nano

SyncDev model directory · Profile updated Nov 19, 2026

Compare with another model

Cosmos3-H-Embodied-Nano is NVIDIA's open 16B surgical world model that does forward dynamics, inverse dynamics and policy from one checkpoint.

Key specifications

Specification score
19.8
License
NVIDIA Open Model License

The specification score summarizes recorded features and limits. It is not a benchmark result.

Pricing conditions: Open weights; no hosted API price. GPU time is the cost, so confirm throughput on your own hardware.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Image input
  • Video input
  • Text input

Produces

  • Video output

Capabilities

VisionOpen weights

What is Cosmos3-H-Embodied-Nano?

Three jobs, one set of weights. Cosmos3-H-Embodied-Nano can turn a first frame and a list of robot actions into surgical video, read actions back out of a video, or act as a policy that predicts both the next actions and the video they will cause. Which job it does depends only on which tokens you leave clean and which you ask it to denoise.

NVIDIA released it on 20 October 2026 for surgical robotics research.

Three modes, mixed in training

  • Forward dynamics: first frame plus actions in, future video out.
  • Inverse dynamics: video in, actions out, useful for labelling unannotated surgical footage.
  • Policy: first frame, a task instruction and action history in, actions and video out.

Training mixed them 0.4 / 0.2 / 0.4 across 50,000 iterations of post-training on the Open-H-Embodiment surgical data: 15 dataset families and 55 dataset roots. JHU dVRK makes up roughly a third of the mixture and CMR Versius clinical procedures about a fifth. Eleven of the fifteen families are public.

Specification

Base Cosmos3-Nano (Qwen3-VL-8B reasoning tower + diffusion generation expert)
Parameters 16B
Video tokenizer Wan2.2
Resolution tier 256p: 320x192, 320x256 or 256x256
Clip 2.4 s, 24 frames and 24 action steps at 10 Hz
Action history Up to 16 past steps
Precision BF16 only tested
Licence NVIDIA Open Model License

Actions are end-effector poses relative to the current pose plus a gripper value, normalised per dataset family. An embodiment ID picks the right action projections.

Compared with what came before

Its predecessor, Cosmos-H-Surgical-Simulator, was a 2B forward-dynamics-only model on Cosmos-Predict2.5. Moving to the Cosmos3 Mixture-of-Transformers and learning both directions of the world–action relationship is the real change. It is also the initialisation for the real-time image-to-video surgical simulator Cosmos3-H-Dreams.

Getting started

Code lives in a plugin, cosmos3_h_embodied, for the unmodified cosmos-framework runtime. It covers Open-H data loading, the training recipe, evaluation and an HTTP policy server that takes a frame, caption and up to 16 past actions and returns 24 actions with predicted video. Weights ship as a BF16 safetensors file and convert to a distributed checkpoint with a bundled tool.

To judge it for your lab, start with inverse dynamics. Label a dozen of your own recorded episodes and compare against the telemetry. It is the cheapest check of whether the model understands your robot before you trust it as a simulator or policy. Then fine-tune on LeRobot-format data from your own embodiment.

More NVIDIA physical-AI models, such as Alpamayo 2 Edge, are on the provider page.

How to evaluate Cosmos3-H-Embodied-Nano

Compare the model on the work you need to complete.

Tasks and inputs

Cosmos3-H-Embodied-Nano is listed for Image-to-video, Video-text-to-text. It accepts image, video, text and produces video. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Choose duration, resolution and frame rate together. Compare the cost of an accepted clip, including repeat generations.

Choosing a deployment

The weights are downloadable under NVIDIA Open Model License. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

Forward dynamics (actions to video), inverse dynamics (video to actions) and policy (frame, text and history to actions plus video) from a single checkpoint.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

tbc

TBC mischief 1

TBC mischief 1 is an image-to-video model built on Wan 2.2, sold on fal and OpenRouter at about half a cent per second for 3-5 s clips at up to 720p.

— context

NVIDIA

Cosmos3-H-Dreams

Cosmos3-H-Dreams is NVIDIA's open real-time surgical world model: it streams Versius robot video from one endoscope frame and two-arm commands.

— context

NVIDIA

Cosmos3-Nano-Sim-Bimanual

Cosmos3-Nano-Sim-Bimanual is NVIDIA's open 16B world model that simulates video of a two-armed AgiBot robot as it moves its camera and manipulates.

— context

noiz-ai

Helix World

Helix World (preview) is an open real-time world model from Noiz AI and HKUST that streams explorable 768x512 video with stereo sound that follows the camera.

— context

Alibaba

Wan-Animate-2

Wan-Animate-2 is Alibaba Wan's open 14B character-animation model that drives a character from video directly, with text-controlled camera angle.

— context

Z.ai / Zhipu AI

SCAIL-2

SCAIL-2 is Zhipu's open character-animation model that drives a reference character from any video, with no pose skeletons or masks. MIT licence.

— context