Nvidia on Oct. 15 unveiled a new open model designed to help companies build a range of AI agents.

Nvidia introduced Nemotron 3.5 Super VL and Nemotron 3.5 Volt for local and edge devices.

A Wall Street Journal report said Nvidia's move is aimed at continuing innovation in open-weight, open-source models and is expected to give momentum to the U.S. open model camp.

Kari Briski (카리 브리스키), vice president of enterprise generative AI software at Nvidia, said the company will continue to release newer and better open models. She called it a message that developers, partners and enterprises can trust and use Nvidia.

Briski said public agencies and financial institutions are reluctant to use China-made models. She said interest is growing because Nemotron is U.S.-made and its training data can be examined. She also highlighted that companies have generally become accustomed to customizing and modifying open models.

Nemotron 3.5 Super VL, with 120 billion parameters, is suitable for developing always-on AI agents.

Nvidia said Nemotron 3.5 Super VL delivers performance similar to its largest reasoning model while being one-fifth the size.

The company said it can also quickly handle agent tasks that take a long time. Briski highlighted that it handles not only text but also images and video.

Several companies, including Atlassian, CrowdStrike, Dassault Systemes, Salesforce and Synopsys, further trained the Nemotron 3.5 Super VL model with their own data. Briski said companies can get better results from a model they directly customize than from a top-end frontier model.

Volt 74B-A9B 24GB, a smaller on-device model based on Super VL, processes agent tasks inside devices. It has 74 billion parameters and is based on a hybrid mixture-of-experts structure in which only 9 billion parameters are activated at a time during operation. The model can run on enterprise workstations or PC graphics processing units, allowing tasks such as coding and debugging to be handled without being sent to the cloud.

Nvidia will release a smaller Volt E4B 4GB at the end of October. It is aimed at the edge for GeForce RTX, Jetson and RTX·GTX Spark, and supports running physical AI and interactive agents within embedded interfaces such as wearables, home devices, point-of-sale terminals, vehicle dashboards and device assist functions.

It can also be used for interactive characters, conversational interfaces, and localized or personalized smart user interfaces.