Nvidia is working on Nemotron 4, a new family of open models designed to compete with the world’s most powerful open-source models. According to those involved, the largest model will have at least 1 trillion parameters. A release date has not yet been set; the final training round is still underway.
The Information reports that Nvidia is working on Nemotron 4, citing employees involved in the project. The largest model in that family is expected to have at least 1 trillion parameters. According to employees, the model could be ready as early as late fall, but the final training has not yet been completed.
Nvidia itself referred to previous statements that it is working on Nemotron 4 but did not confirm the details reported by The Information. However, Kari Briski, vice president of generative AI, issued a statement. “Nvidia is investing in Nemotron because we believe that every company and every country needs accessible, open-source frontier models to strengthen safety and security, accelerate innovation, and provide a foundation they can rely on from generation to generation,” said Briski.
That 1 trillion is a significant leap. The publicly released Nemotron-4 family now has 340 billion parameters and was released under the Nvidia Open Model License, which permits distribution, modification, and commercial use. More recently, Nemotron 3 Ultra was released, an open-weight MoE model with 550 billion parameters, 55 billion of which are active per token, designed for agentic AI and long-term reasoning tasks.
Coalition and open letter
Last month, Nvidia joined forces with other companies to form a coalition aimed at developing and sharing tools for AI safety and cybersecurity. In addition, the company signed an open letter alongside Microsoft and other parties, expressing support for open-weight models to prevent innovation from shifting overseas.
On Tuesday, Nvidia also unveiled Nemotron 3.5 Lightning, designed for tasks including code review, tool usage, monitoring security alerts, and answering billing questions. The company also released NeMo Switchyard, an open-source library for model routing that automatically directs AI tasks to the most suitable model.
Tip: Nvidia establishes Open Secure AI alliance following Hugging Face incident