NVIDIA Cosmos 3 (Super / Nano / Edge)
A family of omnimodal world foundation models aimed at physical AI - robotics, autonomous vehicles and simulation - rather than creative filmmaking.
- Resolution
- 720p
- Aspect ratios
- 16:9, 4:3, 1:1, 3:4
- Weights
- Downloadable
- API
- Yes
- Commercial use
- Vendor states commercial use is permitted
Cosmos 3 ships as Super (64B), Nano (16B) and Edge (4B), with a Reasoner surface for text output and a Generator surface producing images, video, audio and action sequences. Generation supports 256p/480p/720p, 10/16/24/30 fps, and 5 to 300 frames.
Cosmos 3 models are released under the OpenMDW-1.1 License, with custom licensing available via cosmos-license@nvidia.com.
Specification
- What it takes in
- text, image, video, audio
- What it makes
- video, image, audio, text
- Maximum duration
- not established
- Maximum resolution
- 720p
- Aspect ratios
- 16:9, 4:3, 1:1, 3:4, 9:16
- Controls
- text prompt, image or video conditioning, resolution selection, fps selection (10/16/24/30), frame count (5-300), aspect ratio, action-sequence conditioning, 4-step distilled inference, guardrail system
- Weights
- Downloadable
- API
- Yes
- Pricing
- Free open weights, self-hostable. VRAM depends heavily on variant - Edge (4B) targets on-device and edge deployment while Super (64B) needs datacenter GPUs; the repository does not publish a single VRAM table, so no figure is asserted here. NVIDIA also exposes Cosmos models through build.nvidia.com / NIM microservices as a hosted option.
- Released or updated
- 2026-05-31
Maximum duration is not established here. That means no page we could retrieve stated it, not that the model lacks one.
Strengths
- Explicit commercial permission and no revenue or MAU thresholds
- Widest documented control surface: selectable fps, frame count and five aspect ratios
- Three size tiers from 4B edge to 64B datacenter
- Distilled 4-step variants with large claimed speedups
- Strong physics grounding for simulation and robotics
Weaknesses
- Built for physical AI, not cinematic output; aesthetics lag creative-focused models
- 720p ceiling
- Safety-guardrail bypass voids the license
- License lineage changed between generations (NVIDIA Open Model License to OpenMDW-1.1), so check the exact checkpoint
- No consolidated VRAM guidance
Best for
- Robotics, autonomous-driving and simulation teams needing synthetic video and world models with clear commercial terms.
Where the sources disagree
Recorded rather than resolved. Picking the tidier number would hide the fact that the vendor's own pages do not agree.
- License differs by generation: Cosmos-Predict2.5-2B (Oct 2025) is documented as NVIDIA Open Model License, while the Cosmos 3 README (May 2026) states OpenMDW-1.1. Individual Cosmos 3 Hugging Face model pages returned HTTP 401 during this pass, so the OpenMDW-1.1 claim rests on the GitHub README only.
What to understand before using it
Explainers naming the mechanisms this model's own documentation mentions.
See it from the vendor
Shotcraft hosts no sample output. Judging a model means watching its own showreel with your own eyes, and a clip re-hosted here would be a clip we chose.
Nearby in the directory
Adobe Firefly Video Model
The Adobe Firefly Video Model generates video from a text prompt, optionally guided by an image used as the first or final frame. Adobe's Firefly Serv
Video generation
Allegro
A small text-to-video model comprising a 2.8B VideoDiT and a 175M VideoVAE, released under Apache-2.0 in October 2024. It generates 88-frame clips at
Video generation
Amazon Nova Reel 1.1
Amazon Nova Reel generates video from text prompts or from text plus an image, and is available only through Amazon Bedrock. AWS documents output at 1
Video generation