Nvidia released Nemotron 3 Ultra, a new open model (developer.nvidia.com)

Nvidia has announced the release of the Nemotron 3 Ultra, an advanced open model aimed at enhancing the efficiency and effectiveness of long-running AI agents. This 550 billion-parameter Mixture-of-Experts model leverages 55 billion active parameters for improved orchestration and complex planning, allowing it to manage multi-turn workflows while significantly reducing costs—up to 30% according to the company. It excels in handling intricate tasks that require deep reasoning, such as coding, synthesizing information, and verifying designs across extensive constraints. With a throughput up to five times higher than other models in its class, Nemotron 3 Ultra is designed to overcome challenges related to escalating token counts in prolonged agent interactions. The model introduces several technical innovations, including a hybrid Mamba transformer for better sequence efficiency and a new training method called Multi-Teacher On-Policy Distillation (MOPD), which enables continuous improvement through co-evolution with specialized teacher models. It also incorporates the NVFP4 precision check that enhances throughput across different NVIDIA GPU architectures. Additionally, Nemotron 3 Ultra supports a variety of fine-tuning methods and integrates seamlessly with leading agent frameworks, making it an accessible tool for developers aiming to deploy robust AI solutions across various domains. With its commitment to open-source principles and comprehensive support for enterprise applications, this release promises to significantly advance the landscape of AI workflows.