NVIDIA Releases Cosmos3-Nano: 16B-Parameter Omnimodal World Model for Physical AI with 256K Token Context
NVIDIA has released Cosmos3-Nano, a 16-billion parameter omnimodal world model capable of generating video, audio, images, and robot action commands from combinations of text, image, video, and action trajectory inputs. The model supports a 256K token context window and is designed for Physical AI applications including robotics, autonomous vehicles, and smart manufacturing environments.
NVIDIA Releases Cosmos3-Nano: 16B-Parameter Omnimodal World Model for Physical AI
NVIDIA has released Cosmos3-Nano, a 16-billion parameter omnimodal world model that generates video, audio, images, and robot action commands from multimodal inputs. The model is part of the Cosmos3 collection and supports a 256K token context window for reasoning tasks.
Technical Specifications
Cosmos3-Nano uses a Mixture-of-Transformers (MoT) architecture combining an autoregressive transformer for discrete token generation with a diffusion transformer for continuous multimodal generation. The model accepts inputs in five modalities: text, images, video (with or without audio), and action trajectories.
Input specifications:
- Text: up to 256K tokens
- Images: 256p, 480p, and 720p at aspect ratios 16:9, 4:3, 1:1, 3:4, 9:16
- Video: same resolutions/aspect ratios, maximum 5 frames for input
- Audio: 48 kHz stereo, maximum 0.5 seconds
- Action trajectories: 16-400 frames
Output generation supports video from 5 to 400 frames (189 frames default), with audio encoded in AAC format at 48 kHz stereo. The model generates outputs at resolutions matching input specifications.
Model Variants
NVIDIA released multiple Cosmos3 variants simultaneously:
- Cosmos3-Nano: 16B parameters for general omnimodal tasks
- Cosmos3-Super: 64B parameters for enhanced performance
- Cosmos3-Nano-Policy-DROID: 16B parameters specialized for robot manipulation
- Cosmos3-Super-Image2Video: 64B parameters for image-to-video generation
- Cosmos3-Super-Text2Image: 64B parameters for text-to-image generation
Hardware and Deployment
The models require NVIDIA GPU-accelerated systems running on Ampere, Hopper, or Blackwell architectures. Only BF16 precision is officially supported. Runtime integration is available through PyTorch, vLLM-Omni, and Hugging Face Diffusers.
Cosmos3-Nano supports robot action generation for 10 different embodiments, including Franka Panda arms, WidowX 250, and various industrial robots. Action outputs are embodiment-specific, ranging from 9D to 57D depending on the robot platform.
Licensing and Availability
The model is released under the OpenMDW1.1 license for both commercial and non-commercial use. NVIDIA published the model collection on Hugging Face and GitHub on May 31, 2025, with global deployment availability.
Pricing information has not been disclosed.
What This Means
Cosmos3-Nano represents NVIDIA's entry into omnimodal foundation models specifically designed for embodied AI. The 256K context window for reasoning tasks and native support for action trajectory generation distinguishes it from general-purpose multimodal models. The architectural choice to combine autoregressive and diffusion transformers allows different generation mechanisms for discrete (text) versus continuous (video, audio, actions) modalities. With 10 supported robot embodiments at launch, NVIDIA is positioning Cosmos3 as infrastructure for physical AI development rather than a general-purpose model, directly targeting robotics researchers and autonomous system developers who need unified world modeling capabilities.
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.
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.
OpenBMB's MiniCPM-V 4.7 35B-A3B MoE claims 13+ tokens/s on iPhone 17 Pro Max in 3 GB
OpenBMB has released MiniCPM-V 4.7 35B-A3B, the first mixture-of-experts model in the MiniCPM-V series, with open weights on Hugging Face. The company claims it runs at 13+ tokens/s using 3 GB of memory on an iPhone 17 Pro Max, and that it beats Qwen3.6 35B-A3B on overall score in less inference time in Instruct mode.
Comments
Loading...