Nvidia on Monday introduced the latest version of its open model, aimed at helping enterprises build and scale increasingly complex agentic AI applications.
The Nemotron 3 is available in the same three sizes as its predecessor: Nano, Super and Ultra. The models are built on hybrid mix of experts architecture For developers who want to develop and deploy multi-agent systems and solve problems related to transparency and efficiency, Nvidia said.
Nvidia describes Nemotron 3 Super as a high-accuracy logic model with 100 billion parameters and 10 billion active multi-agent applications. Nemotron 3 Ultra is a large-scale reasoning model designed for complex AI applications.
The vendor said the Nemotron 3 Nano is a 30 billion model targeted for highly efficient operations. In addition to the models, Nvidia is also releasing three trillion tokens of new Nemotron pre-training, post-training, and reinforcement learning datasets. The vendor also released the NeMo Gym and NeMo RL open source libraries, which provide a training environment and a post-training foundation. nemotron model,
According to Lian Jae Su, an analyst at Omdia, a division of Informa TechTarget, while the Nemotron 3 is an impressive model, it is not revolutionary or earth-shattering.
“This is an iteration of the previous Nemotron model,” Su said. He said that these models are popular among the developer community. “There is a validity behind the reason that Nvidia maintains the Nemotron model. (With each iteration, it gets smarter.”)
Multi-Agent Systems and Challenges
With this set of models, Nvidia also focuses on another challenge in the AI market: scaling multi-agent systems. For example, Nemotron 3 Nano provides the highest token throughput per second multi-agent systemAccording to Nvidia, this enables the model to remember more and better execute multiple steps in deploying multi-agent systems.
Part of the current challenge with multi-agent systems is that there aren’t many approaches to deploying them, Su said, because they are difficult to implement.
“The complexity behind a multi-agent system lies in the ability for a model to break down different tasks, understand the requirements of each task, and then be able to assign the right task,” he said. Su said that instead of having a larger model perform each task, an emerging approach is to have smaller models perform a single task, with the larger model accountable for overall instruction.
In addition to the challenge enterprises face when deploying multi-agent systems, they also face a specific challenge with Nemotron. Given that Nemotron is open source, models often do not have the support required for enterprise-grade security and governance,
“Nemotron has a lot of the guardrails and firewalls built into the same models that Nvidia introduced,” Su said, referencing previous Nemotron models. “Many enterprises felt it was inadequate.” He said most Nemotron enterprise users create an additional security layer on top of the model. But adding security after the fact requires an enterprise to have a certain amount of expertise in-house.
“Unless you have a powerful internal capability, you don’t understand how to system engineer an excellent multi-agent system, you probably won’t blindly go with the open source model,” Su said. “You are better off getting a more secure model from a verified vendor like Anthropic or OpenAI.”
