Nvidia has released NeMo microservices into general availability with version 25.4, pivoting its profile from a modular toolkit for creating custom generative AI models to emphasizing it as a platform for building AI agents at scale. As AI agents have become an in-demand commodity, Nvidia is leveraging the fact that NeMo’s capabilities seem purpose built to help them grow and thrive. Built around the Kubernetes open-source container management system, NeMo microservices are offered as “an end-to-end developer platform for creating state-of-the-art agentic AI systems,” according to Nvidia.

The agents created by NeMo are continuously optimized using “data flywheels,” a feedback loop where data collected from corporate processes, interactions and user preference is re-applied iteratively — also known as reinforcement learning (RL) — improving the data and outcomes. Or, as the case may be, fine-tuning the agents, making them smarter and more efficient.

AI agents are autonomous digital assistants characterized as non-human co-workers. “Right now enterprises are beginning to build complex multi-agent systems where hundreds of expert agents collaborate to achieve unified goals while working alongside human teammates,” SiliconANGLE writes.

“Maintaining and improving the models that power AI agents in production requires three types of data: inference data to gather insights and adapt to evolving data patterns, up-to-date business data to provide intelligence, and user feedback data to advise if the model and application are performing as expected,” according to an Nvidia blog post. “NeMo microservices help developers tap into these three data types.”

The Wall Street Journal reports Nvidia “is betting on open-source, or open-weight, AI technologies” — like Meta Platforms’ Llama, Google’s Gemma, Microsoft Phi-2 and Mistral AI — “because they tend to offer businesses more flexibility and control than proprietary models.”

“Nvidia’s Llama Nemotron Ultra, currently ranking as the top open model on scientific reasoning, coding and complex math benchmarks, is also accessible through the microservices,” SiliconANGLE explains.