ServiceNow and Nvidia unveil Apriel Nemotron 15B for enterprise agents
ServiceNow and Nvidia just dropped Apriel Nemotron 15B, a compact open-source reasoning model built to run enterprise AI agents at lower latency and inference costs. Delivered as an Nvidia NIM microservice, it pairs with a new joint data flywheel architecture that feeds real enterprise workflow data back into model training.
Apriel Nemotron 15B is a 15-billion-parameter open-source reasoning model built by ServiceNow and Nvidia specifically to power enterprise AI agents.
It packages advanced reasoning capabilities into a lightweight footprint, aiming to slash latency and inference costs while handling complex automated workflows.
What it does
When AI agents execute multi-step workflows without human intervention, they need dedicated reasoning chops to decide how best to proceed at each step.
According to the teams, Apriel Nemotron 15B delivered strong results in benchmark tests for its weight class, proving well-suited for autonomous agentic tasks.
Under the hood, the model was trained using the Nvidia NeMo framework, Nvidia Llama Nemotron Post-Training Dataset, and ServiceNow's domain-specific enterprise data.
How it runs and the data flywheel
The model is packaged to run as an Nvidia NIM microservice, making it straightforward to deploy on Nvidia GPU infrastructure at lower operational cost.
Alongside the model, the two companies introduced a joint data flywheel architecture.
This system integrates ServiceNow Workflow Data Fabric with select Nvidia NeMo microservices, creating a continuous feedback loop that uses live workflow interactions to refine models over time.
It includes built-in enterprise guardrails so companies can fine-tune custom reasoning agents on real operational data without sacrificing privacy or governance.
Why it matters
If you are building autonomous agents for internal business workflows and getting priced out by massive frontier reasoning models, this 15B model hits a sweet spot. Deploying a compact reasoning model via NIM microservices gives you predictable latency and far friendlier unit economics.
It also highlights where enterprise AI is actually heading: smaller, post-trained models tethered to real company data flywheels rather than generic monolithic LLMs. The catch is that the flywheel only works if your internal workflow data is already clean and well-structured.
Sources (1)
- zdnet.com(opens in a new tab)Originalzdnet.com