Nvidia Developing Nemotron 4 With at Least 1 Trillion Parameters to Challenge Top Open-Source AI Models

"Nemotron 4" to have at least 1 trillion parameters

A rare open-source push by a U.S. company draws attention

Nvidia is developing a new family of AI models called Nemotron 4 to compete with the world’s leading open-source models.

The Information reported on Aug. 11, citing Reuters, that the largest Nemotron 4 model under development will have at least 1 trillion parameters.

Nvidia has not set a release date for Nemotron 4 and has yet to complete final training. Employees told the outlet the model could be ready as early as late fall.

Open-source models have drawn more attention this year after China’s Kimi K3 neared the performance of leading AI models from Anthropic and OpenAI. Most U.S. AI labs, however, have not released open-source models.

Interest has also increased after recent hacking incidents involving autonomous AI agents were disclosed, because open-source models do not face restrictions on cybersecurity use.

Last month, Nvidia joined Microsoft and other technology companies to form a coalition to develop and share AI safety and cybersecurity tools. The group also published a letter supporting open-weight models to help keep innovation from moving overseas.

Separately, Nvidia on Aug. 11 unveiled Nemotron 3.5 Lightning, a new product for code review, tool use, security alert monitoring, billing-related questions and other tasks. It also released NeMo-Switchyard, an open-source model-routing library designed to automatically direct AI workloads to the most suitable model.

Open-source models disclose the full development process, including software source code, the data used for training, data preprocessing methods, training scripts and model architecture. Open-weight models, by contrast, do not disclose the full training process or training data. Instead, they release only the weights, or parameters, of an already trained model, allowing anyone to run it or fine-tune it.

Kim Jung-a, guest reporter at Hankyung.com, kja@hankyung.com