ServiceNow and Nvidia Debut Apriel Nemotron 15B, an Open-Source Reasoning Model Built for Faster, Cheaper Agentic AI

Welcome to "The AI Economy," a weekly newsletter by Ken Yeung on how AI is influencing business, work, society, and technology. Subscribe now to stay ahead with expert insights and curated updates—delivered straight to your inbox.

ServiceNow and Nvidia have had a long-standing partnership building generative AI solutions for the enterprise. This week, at ServiceNow’s Knowledge customer conference, the two are introducing the latest fruits of their labor, a new large language model called Apriel Nemotron 15B with reasoning capabilities. The companies believe it performs as well as OpenAI’s o1-mini, Alibaba’s QWQ-32B, and LG’s EXAONE-Deep-32B, but with only half the memory footprint.

Disclosure: I attended ServiceNow's Knowledge conference as a guest of the company, which paid for my flights and hotel. However, no one at ServiceNow dictated what I should write for this post. These words are my own.

This open-source model is designed to help create more intelligent AI agents. It can evaluate relationships, apply rules, and weigh goals to take action. Apriel Nemotron 15B is trained using Nvidia Nemo, Nvidia Llama Nemotron Post-Training Dataset, and ServiceNow’s domain-specific data with Nvidia DGX Cloud on Amazon Web Services.

It’s intended to be used for code assistance and generation, logical reasoning and multi-step tasks, Q&A functionality and information retrieval, and function calling, complex instruction following, and agents. ServiceNow and Nvidia caution against using the model for safety-critical applications that require human oversight.

Perhaps Apriel Nemotron 15B’s signature feature is its ability to provide lower latency and inference costs. In other words, it can provide faster response times and better savings when running this trained AI model (e.g., compute cost, memory usage, energy cost). These things benefit the enterprise because the AI agents powered by Apriel Nemotron 15B will provide real-time quality answers without breaking the bank.

“With this Apriel Nemotron 15B reasoning model, we’re powering intelligent AI agents that can make context-aware decisions, adapt to complex workflows, and deliver personalized outcomes at scale,” ServiceNow’s Executive Vice President of Platform and AI, Jon Sigler, says in a statement.

ServiceNow claims that benchmarks show “promising results” for Nemotron in its model size category, but says it wouldn’t release the evaluations ahead of the announcement per an agreement with Nvidia. However, now that the news has been revealed, both companies say the weights, model, and benchmarks will be publicly available on Hugging Face.

Apriel Nemotron 15B is the latest model from ServiceNow and Nvidia. Last month, it introduced Apriel-5B-Base and Instruct, two small language models. It’s a rarity for ServiceNow, as it hasn’t typically developed AI like this in-house. As Torsten Scholak, the company’s research lead at its Foundation Models Lab, told me in a statement at the time, “While Apriel-5B is not a ServiceNow product and is not deployed in any customer environment, this research helps inform ServiceNow’s broader work in enterprise AI. Insights from this work are being used to improve model efficiency and performance in future initiatives.” And in March, ServiceNow and Nvidia teamed up to pair Nvidia’s Llama Nemotron reasoning models with the ServiceNow platform.

This is also not the first time ServiceNow has partnered with another company to develop open-source AI. In 2023, it worked with Hugging Face on StarCoder 15B, an open-access model.

Apriel Nemotron 15B is estimated to be available in Q2 2025.

However, this LLM isn’t the only release from ServiceNow and Nvidia. Both companies are also announcing a “data flywheel” architecture integrating ServiceNow’s Workflow Data Fabric with Nvidia’s NeMo microservices. “This helps us build AI agents that are contextually aware, deeply personalized, and aligned to the real-time needs of the enterprise,” Sigler attests.

What this means is that the two companies have built a system that can learn from an organization’s internal workflow data while keeping the information secure and under their control. Consequently, this data will help to improve the AI’s reasoning skills and adapt it to the business’s needs.

Featured Image: An AI-generated representation of an artificial intelligence chip inside a computer. Image credit: Adobe Firefly