At the 2026 GTC conference, NVIDIA announced a significant expansion of its open-source model family, aiming to comprehensively accelerate innovation in agents, physics AI, and healthcare. The core of this release is the Nemotron 3 series of fully understanding multimodal models, among which Nemotron 3 Ultra stands out the most. With deep optimization based on the Blackwell architecture, its throughput efficiency has been increased by five times, enabling efficient handling of complex code assistance and enterprise-level workflows.
In addition to the core inference models, NVIDIA also showcased the latest achievements in multimodal interaction: Nemotron 3 Omni seamlessly integrates audio, visual, and language capabilities; Nemotron 3 VoiceChat provides an ultra-low latency real-time conversation experience by connecting speech recognition and synthesis. These models have already received deployment support from industry giants such as CrowdStrike and ServiceNow.
In the field of physical AI and robotics, NVIDIA introduced the world's foundational model Cosmos 3, achieving for the first time the unification of synthetic world generation and action simulation. For embodied intelligence, Isaac GR00T N1.7, designed specifically for humanoid robots, and the autonomous driving inference model Alpamayo 1.5 are now ready for commercial deployment. Huang Renxun also announced the upcoming GR00T N2 at the end of the year, which will double the task success rate in new environments.
Additionally, breakthroughs have also been made in the field of medical research. The BioNeMo platform has added a protein complex generation model, combined with an updated AlphaFold database developed in collaboration with Google DeepMind, which will greatly shorten the drug development cycle. NVIDIA stated that all the above models, datasets, and NIM microservices are now available on GitHub and Hugging Face, allowing developers to easily deploy from the cloud to the edge immediately.
