At the GTC Conference 2026, NVIDIA introduced the Nemotron 3 series—a comprehensive suite of multimodal models. This series comprises three distinct models: Nemotron 3 Ultra, which leverages the Blackwell architecture to achieve a fivefold increase in throughput efficiency; Nemotron 3 Omni, which seamlessly integrates multimodal functionalities; and Nemotron 3 VoiceChat, which facilitates real-time conversational interactions. Additionally, NVIDIA unveiled complementary safety models designed to guarantee the reliability of content outputs. Notably, some of these models are already being deployed by leading enterprises in enterprise-level scenarios. In the realm of physical AI, NVIDIA has released several foundational models and is poised to launch Cosmos 3—the inaugural unified platform for synthetic world generation and motion simulation. Furthermore, two VLA (Vision-Language-Action) inference models are now primed for commercial use, with the next-generation robotic model, GR00T N2, slated for release by year-end. In the field of medical and life sciences research, the BioNeMo platform has expanded its offerings with generative models and has collaborated with relevant institutions to incorporate protein complex predictions into its database. Concurrently, a newly developed GPU-accelerated simulation engine has substantially elevated research efficiency. NVIDIA's open-source models are readily accessible on pertinent platforms, enabling developers to effortlessly deploy them via NVIDIA NIM microservices.