NVIDIAUnited States

Cosmos 3

An AI model: a vision-language-action model (VLA) takes camera images and an instruction and turns them into movements for a robot. This page describes the software, not a robot.

Robotsnot stated

Key figures as stated by the developer

3 of 6 stated · 3 gaps

Model type ⁠A world action model
Controls not stated
Inputs ⁠A camera language audio
Access ⁠A open download
Parameters not stated
Base model not stated

How to read the figures

14kgA30.9lbA: The letter in the grey square after a figure points to the source the figure came from. Robots with several sources carry several letters.
74.5lb: Imperial figures are our own conversion unless the manufacturer states an imperial figure itself. The source letter always belongs to the metric figure — a conversion has no source of its own.

The four states

Versions and notes in the key figures (5)

Model type

Developer wording: "as the backbone for World Action Models (WAMs)".

Controls

Note "Cosmos 3 is an omni-model it can generate across text, image, video, sound and action." "The policy model adapts pre-learned actions to specific tasks, domains, and behaviors at scale."

Inputs

"Generate infinite plausible futures from text, image, video, ambient sound and action input." "Post-train the generalized world foundation model on specialized camera and embodiment data."

Access

"Cosmos 3 is openly available with post-training scripts on GitHub for each modality and module."

Base model

Note "Cosmos 3 is built on Mixture of Transformers architecture."

Full specifications

5 of 12 fields stated

The letter points to the source at the foot of the page. Fields without a figure are dashed: the developer does not state them — that is not the same as zero, and not the same as no.
Field	Value	Source and caveats
Input and output
Model type	`world action model`	A Developer wording: "as the backbone for World Action Models (WAMs)".
Inputs	`camera` `language` `audio`	A "Generate infinite plausible futures from text, image, video, ambient sound and action input." "Post-train the generalized world foundation model on specialized camera and embodiment data."
Controls	not stated	Note "Cosmos 3 is an omni-model it can generate across text, image, video, sound and action." "The policy model adapts pre-learned actions to specific tasks, domains, and behaviors at scale." ⁠A
Degrees of freedom controlled	not stated
Size and speed
Parameters	not stated
Control rate	not stated
Where it runs and what it is built on
Runs where	not stated	Note "Cosmos 3 is optimized for the best performance on NVIDIA hardware." ⁠A
Base model	not stated	Note "Cosmos 3 is built on Mixture of Transformers architecture." ⁠A
Openness
Access	open download	A "Cosmos 3 is openly available with post-training scripts on GitHub for each modality and module."
License	OpenMDW 1.1 (Linux Foundation)	A "Cosmos WFMs are available under the OpenMDW1.1 license from Linux Foundation."
The developer’s own claims
Learns new tasks	needs retraining	A Developer claim: "Post-train Cosmos 3 on embodiment-specific tasks, environments, camera or sensor layouts, and policies".
Works across robots	not stated	Yes only if the developer itself says the model runs on different robot models, for example robots from two manufacturers or two of one manufacturer’s models. Several units of the same robot do not count.

Notes

NVIDIA calls Cosmos 3 the first omni-model with native reasoning, world and action generation.
NVIDIA also offers Cosmos 3 as a controllable, physics-grounded world simulator.
Earlier Cosmos models (Cosmos 2 and 2.5) kept perception and generation as separate models and did not generate actions.

Sources

A Manufacturer product page. www.nvidia.com/en-us/ai/cosmos/ · Retrieved 2026-10-13

One line per source, not per figure: a source has one retrieval date. The letter beside a figure points to the line it came from.