Nvidia targets AI agent market with 9x faster model

Nvidia Corp. is expanding from a hardware provider to a comprehensive platform company with the launch of Nemotron 3 Nano Omni, an open-source model designed to build more efficient enterprise AI agents. The model, which integrates text, vision, and speech capabilities, can deliver up to nine times faster throughput than competing open omni-models, a move that challenges both proprietary models and other open-source alternatives.

"We have embraced NVIDIA Nemotron to reinvent enterprise AI inference for our customers," J.J. Kardwell, CEO of cloud infrastructure company Vultr, said. Vultr, an early adopter, is making the model available on its GPU clusters and through its serverless inference service.

The new model features a 30-billion parameter Mixture-of-Experts (MoE) architecture that only activates 3 billion parameters at any given time, balancing high performance with cost efficiency. By unifying vision and audio encoders within a single framework, it eliminates the need for separate perception modules, reducing latency and cost. The model is designed to run on both high-end consumer hardware and enterprise cloud deployments, and is available as an Nvidia NIM microservice and on platforms like Hugging Face.

The launch positions Nvidia to capture a larger share of the AI value chain, moving beyond selling GPUs to providing the foundational models and tools for agentic AI. This strategy pits Nvidia's "open-and-performant" ecosystem against closed-source leaders and other open-source communities. Early adopters include Palantir Technologies Inc. and Foxconn Technology Group, while companies like Dell Technologies Inc., Oracle Corp., and Infosys Ltd. are currently evaluating the model. The move suggests Nvidia aims to become the go-to provider for not just the "shovels" in the AI gold rush, but the entire construction plan.

The Race for Agentic AI

Nemotron 3 Nano Omni is specifically designed for agentic AI—systems that can understand, reason, and execute complex, multi-step tasks. By training the model on GUI data, Nvidia enables it to comprehend and interact with user interface elements, paving the way for automating office workflows and software operations. "To build useful agents, you can’t wait seconds for a model to interpret a screen," said Gautier Cloix, chief executive of H Company, another early adopter. "By building on Nemotron 3 Nano Omni, our agents can rapidly interpret full HD screen recordings — something that wasn’t practical before." This focus on execution and real-world interaction marks a significant step in the competition to build AI that moves from generating content to performing actions.

An Open Ecosystem Strategy

By releasing Nemotron 3 Nano Omni as an open model, Nvidia is cultivating a developer ecosystem around its hardware. The company is providing not just the model weights but also training data and the NeMo toolkit to foster development. This strategy could attract a broad base of developers and enterprises looking for customizable, high-performance AI solutions without being locked into a closed system. With over 50 million downloads for the Nemotron family in the past year, Nvidia is building a strong foundation. The success of this open, multimodal model could accelerate the adoption of AI agents in enterprises and solidify Nvidia’s central role in the industry's future.

This article is for informational purposes only and does not constitute investment advice.