Fresh off the release of Nemotron 3 Super, a 120-billion-parameter model considered a specialized tool for agentic AI, Nvidia is adding more iterations to the Nemotron family.
NemoClaw, announced Monday at GTC 2026, installs Nvidia's Nemotron models and the newly released Nvidia OpenShell runtime onto the fast-growing open-source AI agent platform OpenClaw in a single command, adding privacy and security guardrails to autonomous AI agents.
How NemoClaw works
NemoClaw runs on any dedicated platform — including GeForce RTX PCs and laptops, RTX PRO workstations, DGX Station and DGX Spark — so teams can run always-on autonomous agents locally without routing sensitive data through the cloud.
NemoClaw uses Nvidia’s Agent Toolkit, which provides a sandbox environment to add privacy controls to agents.
“OpenClaw opened the next frontier of AI to everyone and became the fastest-growing open source project in history,” Huang said. “OpenClaw is the operating system for personal AI. This is the moment the industry has been waiting for — the beginning of a new renaissance in software.”
Nemotron 3 models
Nemotron 3 Ultra, Omni and VoiceChat aim to “extend multimodal intelligence to help deliver specialized, agentic AI.”
Nemotron 3 Ultra “delivers frontier-level intelligence with 5x throughput efficiency with the NVFP4 format on the Nvidia Blackwell platform,” Nvidia said in a press release. During his keynote speech at GTC 2026, Nvidia CEO Jensen Huang said Nemotron 3 Ultra “will be the best model in the business and will help the world build sovereign AI.”
Nemotron 3 Omni uses audio, vision, and language understanding to enable AI agents to extract information from multimodal inputs. Nemotron 3 VoiceChat brings real-time conversations with an AI model that both listens and responds simultaneously. It also “combines automatic speech recognition, LLM processing and text-to-speech capabilities.”
As enterprises increasingly integrate AI agents into their workflows, the need for consistent behavior and security guardrails increases. To this end, Nvidia also released Nemotron safety models and a trustworthy multimodal data retrieval pipeline that detects “unsafe content across text and images, while an agentic retrieval pipeline improves the relevance and accuracy of outputs.”
The Nemotron 3 family of models was first released in December 2025, using a hybrid Mamba-Transformer mixture-of-experts architecture; the Super and Ultra models also use a latent mixture-of-experts design.
Domain-specific and physical AI models
Nvidia is also expanding its domain-specific model portfolio. Its BioNeMo healthcare and life sciences model “is expanding as an open AI development platform” to model, design and simulate biological systems; it also developed Proteina-Complexa to accelerate drug discovery. On the physical AI side, Nvidia unveiled the forthcoming Cosmos 3, a world foundation model that will unify synthetic world generation and physical AI reasoning; the Isaac GR00T N1.7 vision-language-action (VLA) model for humanoid robots; and the Alpamayo 1.5 reasoning VLA model for autonomous vehicles.
These models can run on Nvidia’s NIM microservices platform, allowing enterprises to choose their model deployment.