NVIDIA is developing Nemotron 4 — an open-source AI model with a trillion parameters

NVIDIA is developing a new family of open Nemotron 4 models, which aims to compete with leading open systems worldwide. The flagship version will have no less than one trillion parameters. The company has not announced a release date, as final training is still ongoing. The model could be ready by late autumn. NVIDIA has not confirmed publication details, referring to previous statements about work on Nemotron 4.

Kari Briski, Vice President of Generative AI
The company is investing in Nemotron because it believes that advanced open models are essential for every company and country: they enhance security, accelerate innovation, and provide a foundation that can be built upon from generation to generation.

NVIDIA is among the few major American companies releasing open models. Interest in them has grown this year amid increased AI spending and the emergence of affordable Chinese models that are approaching the capabilities of OpenAI and Anthropic systems. Additional attention to the topic was drawn by a series of disclosed hacks involving autonomous AI agents—open models do not have restrictions on use in cybersecurity. In July, the company formed a coalition with other industry players to develop and share AI security tools. NVIDIA also signed an open letter supporting models with open weights, together with Microsoft and other technology companies.

On the same day, NVIDIA introduced Nemotron 3.5 Lightning—a model for code review, tool usage, security alert monitoring, and billing inquiries. NeMo Switchyard, an open routing library that automatically directs tasks to the most suitable models, was also released.

Source: The Information