NVIDIA Releases Cosmos 3: 64B-Parameter Omnimodal World Model for Physical AI
NVIDIA released Cosmos 3, an omnimodal world foundation model platform for Physical AI spanning robotics, autonomous driving, and industrial environments. The flagship Cosmos3-Super variant contains 64 billion parameters and generates video, images, audio, and action commands from text, image, video, and action trajectory inputs using a Mixture-of-Transformers architecture.
Cosmos 3 Super Image2Video — Quick Specs
NVIDIA Releases Cosmos 3: 64B-Parameter Omnimodal World Model for Physical AI
NVIDIA released Cosmos 3, an omnimodal world foundation model platform designed to accelerate Physical AI development across robotics, autonomous vehicles, and industrial environments. The model collection is available on Hugging Face and GitHub as of May 31, 2026.
Model Specifications
Cosmos 3 comes in five variants with parameter counts ranging from 16B to 64B:
- Cosmos3-Nano: 16 billion parameters
- Cosmos3-Super: 64 billion parameters
- Cosmos3-Nano-Policy-DROID: 16 billion parameters (robotics-specific)
- Cosmos3-Super-Image2Video: 64 billion parameters
- Cosmos3-Super-Text2Image: 64 billion parameters
All models are released under the OpenMDW1.1 license for commercial and non-commercial use.
Technical Architecture
Cosmos 3 uses a Mixture-of-Transformers (MoT) architecture consisting of two complementary transformer towers: an autoregressive transformer for discrete token generation and a diffusion transformer for continuous multimodal generation. Text generates through standard next-token autoregressive decoding, while non-text modalities synthesize through iterative denoising.
Input and Output Capabilities
The models accept multimodal inputs including:
- Text: Up to 256K tokens context window for reasoning tasks
- Images: 256p, 480p, and 720p at aspect ratios 16:9, 4:3, 1:1, 3:4, 9:16
- Video: Up to 5 input frames at the same resolutions
- Audio: 48 kHz stereo with up to 0.5 second duration
- Action trajectories: Compatible with 10 robot embodiments including Franka Panda, UR, Google robot, and WidowX 250
Outputs include video generation from 5 to 400 frames (default 189 frames), images in JPEG format, 48 kHz stereo AAC audio, and robot action sequences.
Robot Platform Support
Cosmos 3 supports action generation for specific robot platforms with dimensionality ranging from 9D (camera motion, UMI) to 57D (egocentric motion). Compatible embodiments include single and dual Franka Panda arms with RobotiQ grippers (10D and 20D), Agibot (29D), and autonomous vehicle control (9D).
Hardware Requirements
According to NVIDIA, the models are optimized for NVIDIA Ampere, Hopper, and Blackwell GPU architectures running Linux. Only BF16 precision is officially supported and tested. The models integrate with PyTorch, vLLM-Omni, and Hugging Face Diffusers runtimes.
What This Means
Cosmos 3 represents NVIDIA's entry into omnimodal world models that bridge multiple AI modalities—vision, language, audio, and robotic control—within a unified architecture. The 256K token context window for reasoning and support for 10 robot embodiments positions it for physical AI applications requiring long-horizon planning. The dual-transformer architecture's separation of discrete and continuous generation mechanisms addresses a core challenge in multimodal modeling. Pricing information has not been disclosed.
Related Articles
Nvidia releases Nemotron 3.5 Super VL: open 120B multimodal MoE with 12B active parameters
Nvidia has released Nemotron 3.5 Super VL, an open multimodal Mixture-of-Experts model with 120 billion total and 12 billion active parameters, built for continuously running enterprise agents. Two compact Bolt models for local PCs and edge devices launch alongside it, though the 4 GB variant ships later this month.
Ollama v0.40.5 adds NVIDIA Nemotron 3.5 Super, a 120B MoE vision-language model with up to 1M-token context
Ollama v0.40.5 adds NVIDIA Nemotron 3.5 Super, a 120B-parameter hybrid Mamba-2/Transformer MoE vision-language model with 12B active parameters and context up to 1M tokens. It also adds a smaller sibling, Nemotron 3.5 Bolt, and fixes a Responses API bug with custom tool calls.
Unbiased releases Pareto 26.9: 262K-context composite model at $2.50/$7.50 per 1M tokens
Unbiased has released Pareto 26.9, a multimodal composite model with a 262K-token context window, priced at $2.50 per 1M input tokens and $7.50 per 1M output tokens. It is listed on OpenRouter with a Oct 13, 2026 release date. A cheaper 1.0M-context preview, Pareto 26.10, is already listed alongside it.
OpenBMB releases MiniCPM-V 4.7 1B, an on-device vision model with 4x/16x visual token compression
OpenBMB has released MiniCPM-V 4.7 1B, a 1B-parameter multimodal model for single-image, multi-image and video understanding on phones, tablets and local hardware. OpenBMB claims it substantially outperforms Qwen3.5 0.8B in overall score and inference efficiency in Instruct mode. Benchmark figures, context window and training cutoff are not given in the available model card text.
Comments
Loading...