Model Reference#

Cosmos 3 is available as Cosmos3-Super (64B), Cosmos3-Nano (16B), and Cosmos3-Edge (4B). The model family also includes task-specific variants: Cosmos3-Super-Text2Image, Cosmos3-Super-Image2Video, their four-step Text2Image and Image2Video variants, plus Cosmos3-Nano-Policy-DROID and Cosmos3-Edge-Policy-DROID. See the Model Matrix for hardware guidance and model selection.

Runtime surfaces#

Surface

Inputs

Outputs

Typical use

Reasoner

Text and vision

Text

World understanding, grounding, physical reasoning, task planning, action forecasting, and decision making

Generator

Text, vision, sound, and action

Vision, sound, and action

World generation, simulation, future prediction, synthetic data generation, and policy learning

Generator and Reasoner are available through different integrations. Use Diffusers or Transformers for Python-first workflows; use vLLM-Omni, vLLM, TensorRT-LLM, or SGLang for OpenAI-compatible serving; or use NIM containers for turnkey deployment. See the Quickstart Guide for the supported integration paths.

Supported generation settings#

Setting

Supported values

Resolution tiers

256p, 480p, 720p (default: 480p)

Aspect ratios

16:9, 4:3, 1:1, 3:4, 9:16 (default: 16:9)

Frame rates

10, 16, 24, and 30 FPS (default: 24)

Frame count

5 to 300 frames (default: 189)

Precision

BF16 tested

Operating system

Linux

GPU architectures

NVIDIA Ampere, Hopper, and Blackwell

Cosmos3-Edge supports 256p and 480p resolutions, 12–30 FPS, and 50–150 frames. It does not currently support video-to-video transfer.

Input and output constraints#

Specification

Value

Input types

Text; text plus image, video, or action

Input formats

Text string; JPG, PNG, JPEG, or WEBP image; MP4 video; JSON action array

Vision conditioning

720p: 1280x720; 480p: 832x480; 256p: 320x192. Video conditioning uses five frames at the matching resolution.

Output types

Image, video, sound, action state, and text

Output formats

JPG image; MP4 video; AAC sound stream muxed into MP4; JSON action values; text string

Prompt guidance

Fewer than 300 words is recommended for world-generation prompts.

Sound output

Stereo AAC at 48 kHz when generated with video; sound is not generated as a standalone output.

Supported action dimensions depend on the embodiment. The source repository lists camera motion and autonomous vehicle (9D), egocentric motion (57D), single-arm robot (10D), dual-arm robot (20D), and humanoid robot (29D) configurations. For the maintained per-integration constraints and examples, refer to the Cosmos repository.