Nvidia announced ongoing development of Nemotron 4, a new open-source AI model family intended to rival leading global open-source AI models. The largest Nemotron 4 model is expected to reach at least 1 trillion parameters, doubling the size of Nemotron 3 Ultra's 500 billion parameters[citation: s1,s2,s4,s6,s7,s9].
As of today, Nemotron 4 has not completed final training nor declared an official release date, but Nvidia projects availability as early as late autumn 2026[citation: s1,s2,s4,s7,s9]. Nvidia’s Vice President of Generative AI, Kari Briski, highlighted the strategic importance of accessible open models, stating, "We believe that every company and every country needs accessible, cutting-edge open models to enhance security and safety, accelerate innovation and provide a foundation that can be relied upon for generations" [1].
In parallel, Nvidia recently released Nemotron 3.5 Lightning, a 30 billion parameter open-source AI model designed to run AI agent tasks such as code review, tool use, security monitoring, and billing queries. Nemotron 3.5 Lightning uses a Mixture of Experts architecture and can run on a single GPU in a typical PC, enabling broad enterprise adoption without licensing fees [2, 3, 4, 5, 6]. This model offers up to four times faster inference speed and improves AI agent task completion by approximately 30% compared to similar models [3, 5].
Nvidia also unveiled NeMo Switchyard, an open-source model routing library that automatically assigns AI tasks to the most appropriate model, improving efficiency and reducing operating costs for users [7, 8, 9, 4, 5, 6]. Nemotron 3.5 Lightning services are available on several platforms, including Hugging Face, ModelScope, OpenRouter, and Nvidia's own build.nvidia.com [3, 5].
Nvidia joined an AI security alliance last month to share AI safety and cybersecurity tools with other companies [7, 8, 9, 4]. CEO Jensen Huang has publicly advocated for open-source AI models, emphasizing that "free AI would benefit hardware and chip demand by expanding AI usage and competition," saying in Mandarin, "免費的AI對硬體來說應該很好。免費的AI對晶片來說應該很好" [2, 6]. In July 2026, Huang urged U.S. government support for open-source AI to prevent innovation from moving overseas [2].
Currently, Chinese open-source AI models like DeepSeek’s V4 and Alibaba’s Qwen exceed the Nemotron 4 target size in parameter count, but Nvidia stresses that a highly capable open model matters more than raw size [7, 6]. The company plans to leverage Nemotron 4’s mix of scale and performance to compete effectively.
Nemotron 4, expected by late autumn 2026, represents Nvidia's push into large-scale open-source AI aimed at enterprises and developers seeking powerful yet accessible AI capabilities [7, 8, 9, 4, 6].
