NVIDIA Launches Nemotron 3 and GR00T N1.7 Open Models for AI Agents and Robotics
Peter Zhang Mar 16, 2026 21:28
NVIDIA unveils expanded open model families at GTC, including Nemotron 3 for agentic AI and GR00T N1.7 for humanoid robots, with major enterprise adoption.
NVIDIA dropped a significant expansion of its open AI model portfolio at GTC on Monday, rolling out new Nemotron 3 variants for enterprise AI agents alongside upgraded robotics models that major manufacturers are already deploying. The chipmaker is betting heavily that open models will cement its dominance beyond GPUs into the software stack powering autonomous systems.
Nemotron 3 Gets Voice, Vision, and Speed
The Nemotron 3 family now spans four distinct models targeting different enterprise use cases. Nemotron 3 Ultra promises 5x throughput efficiency using NVIDIA's NVFP4 format on Blackwell chips—a direct play for companies running coding assistants and complex workflow automation at scale.
More interesting for practical deployment: Nemotron 3 VoiceChat enables simultaneous listening and responding, combining speech recognition, language processing, and text-to-speech in one system. That's the kind of latency reduction that actually matters for customer-facing applications.
CrowdStrike, ServiceNow, Perplexity, and Cursor are already building on Nemotron. Edison Scientific claims its Kosmos platform—powered by Nemotron—compresses months of research into a single day for over 50,000 researchers.
Physical AI Gets Commercial-Ready
The robotics announcements carry more immediate industrial implications. Isaac GR00T N1.7, NVIDIA's vision language action model for humanoids, is now "commercially viable for real-world deployment"—a notable shift from research prototype to production status.
LG Electronics and NEURA are adopting GR00T N1.7 for humanoid robot scaling. Johnson & Johnson MedTech and Toyota Research Institute are using Cosmos for physical AI training.
CEO Jensen Huang previewed GR00T N2, claiming it succeeds at new tasks in new environments more than twice as often as competing VLA models. It currently tops MolmoSpaces and RoboArena benchmarks for generalist robot policies, with availability expected by year-end.
Healthcare Gets Protein Design Tools
NVIDIA partnered with Google DeepMind, EMBL's European Bioinformatics Institute, and Seoul National University to calculate roughly 30 million protein complex predictions, adding 1.7 million high-confidence predictions to the AlphaFold database. Novo Nordisk, Viva Biotech, and Manifold Bio are already testing protein binders generated by the new Proteina-Complexa model.
A new GPU-accelerated simulation engine called nvQSP delivered 77x faster performance than traditional CPU simulations in benchmarks—potentially letting pharma researchers analyze hundreds of dose scenarios before clinical trials.
The Strategic Play
NVIDIA's open model push isn't charity. By providing the software layer that runs optimally on NVIDIA hardware, the company builds switching costs into its ecosystem. Developers building on Nemotron and GR00T become deeply integrated with NVIDIA's stack, from training through deployment via NIM microservices.
The models are available on GitHub, Hugging Face, and build.nvidia.com, with NIM microservice versions for enterprise deployment. Watch for GR00T N2 and Cosmos 3 releases later this year as the next catalysts for the physical AI narrative.
Image source: Shutterstock.jpg)