NVIDIA has announced at the GTC 2026 the expansion of its open model familieswith new LLMs designed to promote agentic AI (Nemotron), as well as physical AI (Cosmos), that intended for autonomous vehicles (Alpamayo), that designed for robotics (Isaac GR00T) and that used in biomedical research (BioNeMo). In addition, it has announced NemoClaw, a stack for the OpenClaw agentic AI platform that provides it with security and brings it closer to the business world.
The family of multimodal models NVIDIA Nemotron It is thus expanded with omnidirectional understanding models that encompass language, vision, voice and security. NVIDIA Nemotron 3 multimodal omnidirectional understanding models power AI agents to deliver natural-style conversations, complete reasoning, and advanced visual capabilities.
Nemotron 3 Ultra brings enhanced performance thanks to the NVFP4 format on the NVIDIA Blackwell platform, and aims to power native AI applications such as scheduling assistants, searches and automation of complex workflows. Nemotron 3 Omni integrates audio, vision and language understanding, allowing AI agents to extract information from videos and documents.
Meanwhile, Nemotron 3 Voice Chat supports real-time conversations where the AI listens and responds at the same time. It combines automatic speech recognition, LLM processing and text-to-speech capabilities in a single system. Nemotron’s security models and recovery channel reinforce reliable multimodal systems by detecting dangerous content in text and images, while an agentic recovery channel improves the relevance and predictability of results.
NVIDIA has also released Nemotron-People, a package of privacy-preserving synthetic datasets based on local census and demographic data. Data sets for France, the United States, Japan, India, Brazil and Singapore are now available.
Among the new models designed to help robots and vehicles perceive, reason and act in the physical world are NVIDIA Cosmos 3a model that unifies synthetic world generation, physical AI reasoning, and action simulation. As to IsaaC GR00T N1.7is an advanced vision, language, and action (VLA) model, designed specifically for humanoids, and is already commercially viable for real-world implementation.
Alpamayo 1.5 is a reasoning VLA model, tasked with driving autonomous vehicle reasoning with navigation guidance, fast conditioning, flexible multi-camera support, and configurable camera parameters. A preview of GR00T N2, a base model for next-generation robots, was also presented at GTC 2026.
NVIDIA BioNeMo It is also being used as an open AI development platform for the healthcare and life sciences sectors. With it, researchers can model, design and simulate large-scale biological systems.
The company has also presented the GPU-accelerated simulation engine nvQSPwith which pharmaceutical researchers will be able to explore more treatment scenarios in computer models before clinical trials begin. It offers much faster performance compared to traditional single-threaded CPU simulations, allowing scientists to analyze hundreds of dose levels and patient subpopulations in significantly less time.
Several of NVIDIA’s open models, data, and frameworks are available on GitHub and Hugging Face, as well as on various cloud infrastructure, inference, and AI platforms. Many of these models are also available as NVIDIA NIM microservices for deployment on any NVIDIA-accelerated infrastructure.
NemoClaw para OpenCloud
NVIDIA has also shown the stack NemoClaw for OpenClaw Agent AI Platformwhich allows users install NVIDIA Nemotron and the NVIDIA OpenShell runtime with a single command. NemoClaw also adds privacy and security controls to make autonomous and self-evolving AI agents, also known as claws, more reliable, scalable and accessible.
NemoClaw uses NVIDIA Agent Toolkit software to optimize OpenClaw with a single command. Install OpenShell to offer open models and an isolated environment that provides data privacy and security to autonomous agents, providing the missing infrastructure layer to provide access to claws to be productive, while applying policy-based security, network, and privacy measures.
NemoClaw is compatible with any scheduling agent. With open agents you can access open models, including NVIDIA Nemotron, that run locally on the user’s dedicated system. With a privacy router, agents can use models that run in the cloud, and the combination of on-premises and cloud models provides a foundation for agents to develop and learn new skills for their tasks, according to defined security and privacy measures.
Always-on agents need dedicated resources to develop software and tools and complete tasks. NemoClaw for OpenClaw can run on any dedicated platform, including PCs and laptops with NVIDIA GeForce RTX or workstations powered by NVIDIA RTX Pro. Also on NVIDIA DGX Station and DGX Spark AI supercomputers, with the aim of offering on-premise computing resources that autonomous agents can operate 24 hours a day.
