Nvidia, ServiceNow engineer open-source model to create AI agents

Nvidia and ServiceNow have created an AI model that can help companies create learning AI agents to automate corporate workloads.

The open-source Apriel model, available generally in the second quarter on HuggingFace, will help create AI agents that can make decisions around IT, human resources and customer-service functions.

[ Related: Agentic AI – Ongoing news and insights ]

“If you look at the foundation models, they’re very big, very slow,” Dorit Zilbershot, ServiceNow’s group vice president of AI experiences and innovation, said in an interview. “This is only a 15-billion-parameters model, it’s highly trained on reasoning. We expect the reasoning to be very, very important.

(By contrast, some of the biggest large language models (LLMs) have more than a trillion parameters and are designed to handle general-purpose tasks.)

The Apriel model should provide faster inferencing at lower cost compared to those larger models, Zilbershot said.

[ Related: More Nvidia news and insights ]

The model was trained on the open-source Nvidia NeMo LLM, the NVIDIA Llama Nemotron Post-Training Dataset, and ServiceNow domain-specific data, the companies said in a joint statement.

AI agents have been closely associated with the concept of “digital labor” and can help workers be more productive and efficient by evaluating and establishing relationships between data in an enterprise. With that understanding, the agents can automate tasks, saving workers time and allowing them to make smarter decisions.

More tools can be attached to the inferencing model. An orchestrator, for instance, can take the role of a “manager” that creates the digital labor based on learnings from Apriel.

“You can think about the orchestrator as the manager that leverages the LLM that can reason and plan and kind of understand which…agents to run when,” Zilbershot said.

Beyond that, there are also tools to establish workflow and surface data to better understand the task at hand and the context to correctly answer questions. The integration of Apriel, the orchestrator and data and workflow tools can yeild a sort of digital CEO for AI agents.

“We look at these three together, the AI the data and actions or the workflows as the key pieces to make AI agents,” Zilbershot said. “It’s where you get that boss mode [for AI agents].”

Agentic AI is still in its early stages with tech leaders remaining cautious about its implementation. Early adopters have highlighted  regulatory and trust concerns, and many customers simply prefer to work with humans over computers.

But ServiceNow and rivals Atlassian and Salesforce are actively integrating AI agents into workflows. Those companies need reasoning models for customers to create and run AI agents. For example, Google’s AI agents run on its Gemini models.

Salesforce CEO Marc Benioff recently talked about the company’s embrace of AI agents, likening the technology to Ironman’s Jarvis. And Atlassian is creating collections of integrated apps designed to make business processes, collaboration, analysis and execution more efficient.

ServiceNow offers its own LLM — NowLLM — and models that include OpenAI’s ChatGPT-4o. The company also offers thousands of AI agents on its platform.

“We had customers go live with our agents less than a week after we went GA. People understand the technology. We built it in a way that it’s fully embedded within ServiceNow,” Zilbershot said. 

While AI agents are already driving productivity mostly at the  department level, things will change within a few of months.

“I would predict that within three to six months, we’ll see those cross-functional agents and you’ll see even higher rates of productivity gains,” Zilbershot said.