NVIDIA has announced several advancements in AI at CES 2025, unveiling new developments that merge the company’s previous successes in synthetic data generation with its focus on autonomous decision-making.
The new releases include the Cosmos World Foundation Model (WFM) platform—an extensive toolkit for creating photoreal, physics-based videos and scenarios—and the Llama Nemotron plus Cosmos Nemotron families, which enable language, vision, and decision-making AI in diverse sectors such as robotics, healthcare, and autonomous vehicles.
Related: Nvidia Reveals RTX 50-Series Blackwell GPUs with DLSS 4 and Twice the Power of the RTX 4090
“We created Cosmos to democratize physical AI and put general robotics in reach of every developer,” said Jensen Huang, NVIDIA’s founder and CEO. “AI agents is the next robotic industry and likely to be a multibillion-dollar opportunity.”
By combining synthetic data generation, vision processing, and advanced language models under one umbrella, NVIDIA aspires to streamline the transition from data creation to fully operational AI systems. This approach follows the success of the Nemotron-4 340B series, which previously addressed shortages of high-quality training data for large language models (LLMs).
Nemotron-4 340B: Establishing a Data-Driven Foundation
In mid-2024, NVIDIA introduced the Nemotron-4 340B models to tackle limited data availability for complex AI applications. These models produced synthetic data at scale, enabling high-level refinement and adaptation for industries like healthcare, finance, and manufacturing.
Nemotron-4 340B offered three variants—Base, Instruct, and Reward. The Instruct models helped developers guide AI outputs through clear directives, while the Reward models scored the generated responses based on parameters such as accuracy and coherence. This iterative feedback mechanism proved valuable for training large language models, speeding up development and improving model reliability.
The Nemotron-4 340B initiative also integrated seamlessly with NVIDIA’s NeMo platform and TensorRT-LLM library, providing users with optimization and flexibility in their AI workflows. The synthetic data generated by Nemotron-4 340B laid the groundwork for NVIDIA’s latest breakthroughs in agentic and physical AI, bridging data curation, model training, and deployment needs.
Llama Nemotron and Cosmos Nemotron: Expanding Agentic AI
NVIDIA’s newest offerings in the Nemotron family—Llama Nemotron and Cosmos Nemotron—move beyond just data generation to power real-time AI agents. Llama Nemotron large language models (LLMs) cater to tasks such as coding, function calling, chat, and mathematical computations, while Cosmos Nemotron vision language models (VLMs) focus on interpreting and responding to visual data in videos, images, and sensor feeds.
“Agentic AI is the next frontier of AI development, and delivering on this opportunity requires full-stack optimization across a system of LLMs to deliver efficient, accurate AI agents,” said Ahmad Al-Dahle, vice president and head of GenAI at Meta, in a statement. “Through our collaboration with Nvidia and our shared commitment to open models, the Nvidia Llama Nemotron family built on Llama can help enterprises quickly create their own custom AI agents.”
This dual-pronged approach incorporates specialized NVIDIA NIM microservices that handle resource-heavy tasks like video search, summarization, and sensor interpretation. By integrating language and visual processing, AI agents can manage a range of applications, from warehouse logistics to medical imaging analysis.
Cosmos World Foundation Models
Alongside the Llama Nemotron and Cosmos Nemotron families, NVIDIA launched the Cosmos World Foundation Model (WFM) platform. This new platform specializes in generating photoreal, physics-based videos and environments for robotics, autonomous vehicles, and general “physical AI” scenarios. Its focus on realistic simulations lowers the costs associated with collecting and testing massive amounts of real-world data.
“The ChatGPT moment for robotics is coming. Like large language models, world foundation models are fundamental to advancing robot and AV development, yet not all developers have the expertise and resources to train their own,” noted Huang in his opening keynote at CES.
Developers can use Cosmos WFMs to create tailored scenarios, adding complexities such as snowy roads for AV systems or congested warehouse floors for robotics testing. These physics-aware datasets can either refine existing models or serve as a standalone training resource. The company has made these models available under an open model license, aiming to broaden access to advanced AI development.
Accelerating Physical AI Through Data and Compute Efficiency
Physical AI remains computationally demanding, requiring high-fidelity data to simulate the real world. Cosmos addresses these challenges by offering an accelerated video processing pipeline, advanced video tokenizers (available under NVIDIA’s open model license, via Hugging Face and GitHub), and the NVIDIA NeMo Curator for data labeling and curation.
This pipeline aims to process vast amounts of video data—up to 20 million hours in 14 days using the NVIDIA Blackwell platform—rather than years of CPU-bound operations.
These efficiency gains help organizations seeking to develop, test, and refine their AI models without being limited by real-world data constraints. Cosmos Tokenizer compresses images and videos, reducing overhead while preserving essential quality for training advanced AI systems. According to NVIDIA, these optimizations pave the way for faster iteration in robotics and autonomous vehicle research.
Industry Adoption
Major players in robotics and automotive technology have shown strong interest in Cosmos. Companies such as 1X, Agile Robots, Agility, Figure AI, Foretellix, Uber, Waabi, and XPENG are among those integrating the new platform into their development pipelines.
For instance, XPENG plans to enhance its humanoid robotics initiatives, while ridesharing giant Uber collaborates with NVIDIA to harness Cosmos for better data curation and scenario generation. “Generative AI will power the future of mobility, requiring both rich data and very powerful compute,” said Dara Khosrowshahi, CEO of Uber. “By working with NVIDIA, we are confident that we can help supercharge the timeline for safe and scalable autonomous driving solutions for the industry.”
Companies like SAP and ServiceNow have similarly embraced NVIDIA’s Nemotron families. “AI agents that collaborate to solve complex tasks across multiple lines of the business will unlock a whole new level of enterprise productivity beyond today’s generative AI scenarios,” said Philipp Herzig, chief AI officer at SAP, in a statement. “Through SAP’s Joule, hundreds of millions of enterprise users will interact with these agents to accomplish their goals faster than ever before.”
NeMo Integration, Open Licensing, and Safety Measures
All Cosmos WFMs and Nemotron models interface with NVIDIA’s NeMo framework, enabling fine-tuning, alignment, and retrieval-augmented generation (RAG). Through NeMo Curator, developers can process large-scale video data, while reinforcement learning from human feedback (RLHF) refines the models to maintain appropriate, context-driven responses.
NVIDIA has released Cosmos under an open model license, encouraging collaboration and customization within the robotics and AV community. The company also noted measures for safe and responsible AI, including watermarking AI-generated content, implementing guardrails to mitigate harmful text or images, and aligning with global AI safety initiatives.
“We are confident that we can help supercharge the timeline for safe and scalable autonomous driving solutions for the industry,” added Khosrowshahi, underscoring a growing emphasis on trustworthy, transparent AI systems.
Toward a Unified AI Ecosystem
By merging the synthetic data-driven approach of Nemotron-4 340B with the new Cosmos WFM platform, NVIDIA sets forth a unified path for AI that spans research, enterprise deployment, and physical automation. Llama Nemotron and Cosmos Nemotron families fill key roles in agentic AI, while Cosmos WFMs address the complexities of robotics and autonomous vehicle development.
From enabling cost-effective data generation to offering specialized microservices for real-time language and vision tasks, NVIDIA’s latest portfolio exemplifies a versatile strategy for AI advancement. As more enterprises, developers, and researchers adopt these models, the trajectory for autonomous systems and intelligent software agents appears ready to accelerate.
