Nvidia CEO Jensen Huang stated final yr that we at the moment are getting into the age of bodily AI. Whereas the corporate continues to supply LLMs for software program use instances, Nvidia is more and more positioning itself as a supplier of AI fashions for absolutely AI-powered techniques — together with agentic AI within the bodily world.
At CES 2026, Nvidia introduced a slate of latest fashions designed to push AI brokers past chat interfaces and into bodily environments.
Nvidia launched Cosmos Purpose 2, the most recent model of its vision-language mannequin designed for embodied reasoning. Cosmos Purpose 1, launched final yr, launched a two-dimensional ontology for embodied reasoning and at present leads Hugging Face’s bodily reasoning for video leaderboard.
Cosmos Purpose 2 builds on the identical ontology whereas giving enterprises extra flexibility to customise purposes and enabling bodily brokers to plan their subsequent actions, just like how software-based brokers purpose by digital workflows.
Nvidia additionally launched a brand new model of Cosmos Switch, a mannequin that lets builders generate coaching simulations for robots.
Different vision-language fashions, comparable to Google’s PaliGemma and Pixtral Massive from Mistral, can course of visible inputs, however not all commercially out there VLMs help reasoning.
“Robotics is at an inflection level. We’re shifting from specialist robots restricted to single duties to generalist specialist techniques,” stated Kari Briski, Nvidia vp for generative AI software program, in a briefing with reporters. She was referring to robots that mix broad foundational information with deep task-specific expertise. “These new robots mix broad basic information with deep proficiency and sophisticated duties.”
She added that Cosmos Purpose 2 “enhances the reasoning capabilities that robots must navigate the unpredictable bodily world.”
Transferring to bodily brokers
Briski famous that Nvidia’s roadmap follows “the identical sample of property throughout all of our open fashions.”
“In constructing specialised AI brokers, a digital workforce, or the bodily embodiment of AI in robots and autonomous automobiles, extra than simply the mannequin is required,” Briski stated. “First, the AI wants the compute sources to coach, simulate the world round it. Knowledge is the gasoline for AI to be taught and enhance and we contribute to the world's largest assortment of open and various datasets, going past simply opening the weights of the fashions. The open libraries and coaching scripts give builders the instruments to purpose-build AI for his or her purposes, and we publish blueprints and examples to assist deploy AI as techniques of fashions.”
The corporate now has open fashions particularly for bodily AI in Cosmos, robotics, with the open-reasoning vision-language-action (VLA) mannequin Gr00t and its Nemotron fashions for agentic AI.
Nvidia is making the case that open fashions throughout completely different branches of AI kind a shared enterprise ecosystem that feeds information, coaching, and reasoning to brokers in each the digital and bodily worlds.
Additions to the Nemotron household
Briski stated Nvidia plans to proceed increasing its open fashions, together with its Nemotron household, past reasoning to incorporate a brand new RAG and embeddings mannequin to make data extra available to brokers. The corporate launched Nemotron 3, the most recent model of its agentic reasoning fashions, in December.
Nvidia introduced three new additions to the Nemotron household: Nemotron Speech, Nemotron RAG and Nemotron Security.
In a weblog put up, Nvidia stated Nemotron Speech delivers “real-time low-latency speech recognition for dwell captions and speech AI purposes” and is 10 occasions sooner than different speech fashions.
Nemotron RAG is technically comprised of two fashions: an embedding mannequin and a rerank mannequin, each of which might perceive photographs to offer extra multimodal insights that information brokers will faucet.
“Nemotron RAG is on prime of what we name the MMTab, or the Large Multilingual Textual content Embedding Benchmark, with sturdy multilingual efficiency whereas utilizing much less computing energy reminiscence, so they’re a great match for techniques that should deal with a number of requests in a short time and with low delay,” Briski stated.
Nemotron Security detects delicate information so AI brokers don’t unintentionally unleash personally identifiable information.
[/gpt3]
