At the 2026 GTC conference held in San Jose, California, on Monday, NVIDIA announced a comprehensive expansion of its open-source model family to drive the next wave of Agentic, physical, and medical AI.
In the field of intelligent agents, NVIDIA launched the Nemotron 3 series of fully understanding multimodal models. Among them, the Nemotron 3 Ultra, based on the Blackwell architecture, offers a 5x improvement in throughput efficiency and is designed specifically for code assistance and complex workflows.
The Nemotron 3 Omni precisely integrates audio, visual, and speech multimodal capabilities; while the Nemotron 3 VoiceChat integrates speech recognition, large models, and speech synthesis to support real-time dialogue.
The accompanying security model further ensures the reliability of content output, and leading companies such as CrowdStrike and ServiceNow have already deployed it in their enterprise applications.
In the field of physical AI, NVIDIA has released several foundational models to accelerate the development of autonomous systems. The upcoming Cosmos 3 is the first foundational model for unified synthetic world generation and motion simulation; Isaac GR00T N1.7 and Alpamayo 1.5 are visual language action (VLA) inference models specifically designed for humanoid robots and autonomous driving, respectively, and are now ready for commercial deployment.
In addition, CEO Jensen Huang also previewed the next-generation robot model GR00T N2, which has doubled the success rate of existing models in new environments and is expected to be launched by the end of the year.
Significant breakthroughs have also been achieved in medical and life science research. According to a blog post cited by Free Tech News, the BioNeMo platform has added the generative model Proteina-Complexa, which has greatly accelerated the discovery of protein drugs.
Meanwhile, NVIDIA, in collaboration with Google DeepMind and other institutions, added 1.7 million high-confidence protein complex predictions to the AlphaFold database. Furthermore, the new GPU-accelerated simulation engine nvQSP is up to 77 times faster than traditional CPUs, enabling scientists to analyze massive amounts of treatment plans in a very short time.
The aforementioned open-source models, datasets, and frameworks are now available on platforms such as GitHub and Hugging Face. Developers can also leverage NVIDIA NIM microservices to achieve secure, scalable, and convenient deployment across various acceleration infrastructures from the edge to the cloud.
