Nvidia's AI agent play is here with new models, orchestration blueprints

The industry push to agent AI continues, with Nvidia announces several new services and modules to facilitate the creation and use of AI agents.

Today, Nvidia launched Nemotron, a family of modules based on it Meta's Llama and received training on the company's methods and databases. The company also announced new AI orchestration plans to manage AI agents. These latest announcements bring Nvidia, a company best known for the hardware that powers the AI ​​generation revolution, to the forefront of AI agent development.

Nemotron comes in three sizes: Nano, Super and Ultra. It also comes in two flavors: the Nemotron Llama for language tasks and the Cosmos Nemotron vision module for physical AI projects. The Nemotron Nano Llama has 4B parameters, the Super 49B parameters and the Ultra 253B parameters.

All three work best for agent tasks including “direction following, chat, call to action, coding and math,” according to the company.

Rev. Lebaredian, VP of Omniverse and simulation technology at Nvidia, said in a briefing with reporters that the three sizes are optimized for different Nvidia computing resources. Nano is for cost-effective low latency applications on PC and peripheral devices, Super is for extreme accuracy and throughput on a single GPU and Ultra is for the highest accuracy at data center scale.

“AI agents are the digital workers who work for us and work with us, so the Nemotron model family is for agent AI,” Lebaredian said.

The Nemotron models are available as hosted APIs on Hugging Face and Nvidia's website. Nvidia said enterprises will be able to access the models through its​​​​ AI Enterprise software platform.

Nvidia is no stranger to basic models. Last year, it was quietly released version of Nemotron, Llama-3.1-Nemotron-70B-Instructthat outperformed similar models from Open AI and Anthropic. It is also released NVLM 1.0a family of multimodal language models.

More support for agents

AI Agents will be a big trend in 2024 when enterprises start exploring how to use agent systems in their workflow. Many believe that momentum will follow this year.

Companies like Sales force, Service now, AWS and Microsoft all have called agents the next wave of gen AI in enterprises. AWS is added multi-agent orchestration to Bedrock, and Salesforce released the Agentforce 2.0giving more agents to their customers.

However, agent workflows still require other infrastructure to function effectively. One such infrastructure revolves around orchestration, or managing multiple producers spanning different systems.

Orchestration plans

Nvidia has also entered the new realm of AI orchestration with its blueprints that guide agents through specific tasks.

The company has partnered with several orchestral companies, among them LangChain, Index of lama, CrewAI, every day and Weight and biasto build blueprints on Nvidia AI Enterprise. Each orchestration framework has its own blueprint developed by Nvidia. For example, CrewAI created a plan for code documentation to ensure that code sources are easy to navigate. LangChain added Nvidia NIM microservices to its structured report generation plan to help agents return Internet searches in various formats.

“Getting multiple agents to work together smoothly or orchestration is key to using agent AI,” Lebaredian said. “These leading AI orchestration companies integrate all Nvidia, NIM, Nemo and Blueprints agent building blocks with their open source agent orchestration platforms.”

Nvidia's new PDF-to-podcast plan aims to compete with that GoogleLM Notebook by converting information from PDF to audio. Another new plan will help build agents to find and summarize videos.

Lebaredian said Blueprints aims to help developers quickly deploy AI agents. To that end, Nvidia unveiled Nvidia Launchables, a platform that allows developers to test, prototype and run blueprints in one click.

Orchestration could be one of the bigger stories in 2025 as enterprises engage in multi-agent production.