NVIDIA does not just make chips. It has spent years building a parallel business in AI software and open models, and at GTC this week the company made clear just how broad that effort has become. The new model releases span four distinct domains: agentic AI for software and enterprise applications, physical AI for robots and autonomous vehicles, healthcare AI for drug discovery and medical research, and multimodal AI that combines language, vision, and voice in a single system. The companies already deploying these models range from cybersecurity firms to pharmaceutical researchers, which gives a sense of how far the ambition here extends.
The Nemotron Family Gets a Major Expansion
The centerpiece of the agentic AI announcements is the Nemotron 3 lineup, which NVIDIA is positioning as the foundation for building AI agents that can handle real conversations, complex reasoning tasks, and visual understanding all at once.
Nemotron 3 Ultra is the highest-capability model in the family, delivering what NVIDIA describes as frontier-level intelligence with five times the throughput efficiency of previous formats when running on the Blackwell platform using the NVFP4 format. It is aimed at demanding applications like coding assistants, intelligent search, and complex workflow automation where performance and speed both matter.
Nemotron 3 Omni takes a different approach, combining audio, vision, and language understanding in a single model. The practical benefit is that an agent built on Omni can extract insights from videos and documents simultaneously, handling the kind of mixed-format information that real-world tasks frequently involve. Nemotron 3 VoiceChat pushes the real-time conversation angle further, combining automatic speech recognition, large language model processing, and text-to-speech in one integrated system so that AI can genuinely listen and respond at the same time rather than processing each step sequentially.
Rounding out the Nemotron announcements are new safety models and a retrieval pipeline designed to make multimodal systems more trustworthy. The safety models detect unsafe content across both text and images, while the agentic retrieval pipeline improves the relevance and accuracy of what agents pull in when they need to look something up.
The adoption list for Nemotron is already substantial. CodeRabbit, CrowdStrike, Cursor, Factory, Perplexity, and ServiceNow are all deploying Nemotron models for agentic applications. LangChain has integrated Nemotron alongside other NVIDIA Agent Toolkit software into its agent development platform. Edison Scientific is using Nemotron as a core component of Kosmos, an autonomous AI scientist used by more than 50,000 researchers that can compress months of research into a single day by running hundreds of research tasks in parallel.
NVIDIA has also released Nemotron-Personas, a collection of privacy-preserving synthetic datasets grounded in local census and demographic data, designed to help developers build AI that understands and reflects local cultures and languages. The France dataset, developed with Pleias, is available now, joining existing datasets for the United States, Japan, India, Brazil, and Singapore. Developers in countries including Singapore, Poland, Indonesia, France, Vietnam, and Malaysia are already using Nemotron frameworks to build sovereign AI models that serve users in their native languages.
Physical AI Gets Three New Foundation Models
The second major thread of the announcement covers physical AI, meaning the models that help robots and autonomous vehicles perceive and interact with the real world. This is an area NVIDIA has been investing in heavily, and the new releases represent meaningful advances in what is commercially deployable right now.
NVIDIA Isaac GR00T N1.7 is an open reasoning vision language action model built specifically for humanoid robots. The key update here is that NVIDIA is now describing it as commercially viable for real-world deployment, a distinction that matters because earlier versions were primarily research tools. Humanoid, LG Electronics, NEURA, and Noble Machines are already adopting GR00T N1.7 to scale humanoid robot deployment, which suggests the commercial viability claim has real backing.
NVIDIA Alpamayo 1.5 is a reasoning vision language action model aimed at autonomous vehicles. It adds navigation guidance, prompt conditioning, flexible multi-camera support, and configurable camera parameters to the autonomous vehicle development stack, giving engineers more tools to customize how their systems perceive and respond to driving environments.
NVIDIA Cosmos 3 is described as the first world foundation model to unify synthetic world generation, physical AI reasoning, and action simulation in a single model. It is expected to arrive soon and is positioned as a tool for helping physical AI systems handle complex, unpredictable real-world environments by training on simulated scenarios generated by the model itself. HCLTech, Johnson and Johnson MedTech, Milestone Systems, mimic robotics, Skild AI, Tulip, and the Toyota Research Institute are all using NVIDIA Cosmos to accelerate physical AI training and video analytics.
Jensen Huang also used his GTC keynote to preview GR00T N2, a next-generation robot foundation model based on research called DreamZero. Built on a new world action model architecture, GR00T N2 helps robots succeed at new tasks in new environments more than twice as often as leading vision language action models, according to NVIDIA. It currently holds the top position on both MolmoSpaces and RoboArena for generalist robot policies and is expected to be available by the end of the year.
Healthcare AI Is Getting Serious Computational Muscle
The healthcare and life sciences portion of the announcement is where some of the most concrete and scientifically significant work is happening. NVIDIA is expanding BioNeMo as an open development platform for healthcare and life sciences research, and the new additions cover protein design, drug discovery, and pharmaceutical simulation.
Proteina-Complexa is a generative model for protein binder design that accelerates structure-based drug discovery. A protein binder is a molecule designed to attach to a specific target protein, which is a fundamental step in developing many types of drugs. Novo Nordisk, Viva Biotech, and Manifold Bio are all using Proteina-Complexa to design protein binders and have already experimentally tested the designs the model generates, which is a meaningful signal that the model is producing outputs worth taking into the lab.
NVIDIA has also collaborated with EMBL’s European Bioinformatics Institute, Google DeepMind, and Seoul National University to significantly expand the AlphaFold Protein Structure Database. The collaboration calculated approximately 30 million protein complex predictions and added 1.7 million high-confidence predictions to the existing database. The AlphaFold database is already one of the most important resources in structural biology, and expanding it at this scale could meaningfully accelerate the discovery of new drug targets and improve the understanding of disease biology.
The third healthcare release is nvQSP, a GPU-accelerated simulation engine for pharmaceutical researchers. It enables scientists to explore far more treatment scenarios in computer models before committing to clinical trials, which are expensive, time-consuming, and ethically constrained in terms of how many variables can be tested at once. In benchmark testing, nvQSP delivered up to 77 times faster performance compared with traditional single-threaded CPU simulations, allowing researchers to analyze hundreds of dose levels and patient subpopulations in the time it previously took to simulate just a few scenarios. That is the kind of speedup that can genuinely change how drug development workflows are structured.
Why the Open Model Strategy Matters
It would be easy to read this announcement as a list of products and move on, but the strategic logic behind it is worth understanding. NVIDIA is not just releasing models for researchers to experiment with. It is building a portfolio that positions NVIDIA software as foundational infrastructure across every major application domain for AI, and it is doing so through open models that developers can actually use, modify, and build on.
The breadth of the adoption already in place makes this concrete. Cybersecurity companies are using Nemotron to build smarter threat detection agents. Pharmaceutical companies are using BioNeMo models to accelerate drug discovery pipelines. Robotics companies are using GR00T to deploy humanoids. Automotive companies are using Alpamayo to develop autonomous driving systems. That range of applications across a single model release cycle is unusual, and it reflects how deliberately NVIDIA has structured its open model portfolio around the verticals where AI is having the most immediate real-world impact.
Kari Briski, vice president of generative AI software at NVIDIA, framed the goal as extending intelligence beyond language, enabling developers worldwide to build intelligent agents and power breakthroughs across digital and physical industries. That description captures the ambition accurately. The question of whether NVIDIA’s open model strategy compounds into something as foundational as its hardware business is one of the more interesting long-term stories in AI to watch right now.
Where to Get the Models
Select NVIDIA open models, data, and frameworks are available on GitHub and Hugging Face, across a range of cloud, inference, and AI infrastructure platforms, and on build.nvidia.com. Many are also available as NVIDIA NIM microservices for secure, scalable deployment on any NVIDIA-accelerated infrastructure from the edge to the cloud.