NVIDIA announced Nemotron 3 Ultra, a 550-billion-parameter mixture-of-experts open model aimed at long-running enterprise AI agents, alongside software for building and running autonomous workflows. The company says the model will be available through Hugging Face, ModelScope, OpenRouter, build.nvidia.com and cloud...
NVIDIA announced Nemotron 3 Ultra, a 550-billion-parameter mixture-of-experts open model designed for long-running enterprise AI agents.
In a June 1, 2026 announcement, NVIDIA said Nemotron 3 Ultra is intended to support autonomous enterprise AI agents that can work across complex business and technical workflows. According to NVIDIA’s investor relations release, the company claims the model can deliver up to 5x faster inference and up to 30% lower cost compared with “open frontier models in its class.”
NVIDIA said availability is expected on June 4 through Hugging Face, ModelScope, OpenRouter, build.nvidia.com and cloud partners. The company described Nemotron 3 Ultra as an open model, positioning it for enterprises and developers that want to build AI systems with more control over deployment and integration.
The announcement fits into NVIDIA’s broader push to provide not just chips, but also model, runtime and software components for companies building AI applications. NVIDIA Newsroom said the launch includes Agent Toolkit software, NemoClaw blueprints, Nemotron models, the OpenShell runtime and CUDA-X agent skills for building autonomous enterprise AI agents.
NVIDIA named several enterprise software and infrastructure companies in connection with the announcement, including Cadence, Siemens, Synopsys, CrowdStrike, Palantir, Microsoft, Canonical and Red Hat, according to NVIDIA Newsroom.
The partner list suggests NVIDIA is aiming the technology at several enterprise categories: engineering and design software, cybersecurity, data platforms, cloud services and Linux infrastructure. NVIDIA did not frame the announcement as a single product launch only; instead, it presented Nemotron 3 Ultra as part of a software stack for building agents that can operate across enterprise environments.
Cadence separately said its ChipStack AI Super Agent uses NVIDIA Nemotron models and runs within NVIDIA OpenShell. In its press release, Cadence described ChipStack as a “fully autonomous virtual engineer” for chip design and verification workflows. The company said the system is intended to support autonomous tasks in semiconductor design, an area where verification, optimization and iteration can be time-consuming.
NVIDIA’s claims center on performance, cost and deployment options. The company’s investor relations release says Nemotron 3 Ultra is a 550B-parameter MoE model, a design that typically activates only part of the model for a given task rather than using all parameters at once. NVIDIA’s stated performance and cost comparisons are company claims and were not independently verified in the provided sources.
The company is also emphasizing runtime and tooling. NVIDIA Newsroom said OpenShell, Agent Toolkit software, NemoClaw blueprints and CUDA-X agent skills are part of the broader offering. Those components are positioned to help developers build autonomous systems that can use tools, interact with software environments and carry out longer tasks than a single prompt-and-response exchange.
For enterprises, the practical question will be whether these components can be integrated into existing systems with acceptable reliability, security and governance. The source materials describe partner adoption and planned availability, but they do not provide independent benchmarks, customer case studies with measured outcomes, or detailed pricing.
The announcement shows NVIDIA continuing to expand from AI hardware into enterprise AI software infrastructure. By releasing Nemotron 3 Ultra through widely used model platforms such as Hugging Face and ModelScope, as well as NVIDIA’s own build.nvidia.com and cloud partners, the company is trying to make the model accessible across multiple developer channels.
The Cadence example also shows where NVIDIA expects autonomous AI agents to be used first: specialized, high-value workflows where domain software already exists and where automation may be easier to measure. Chip design and verification are complex engineering tasks, and Cadence’s use of Nemotron models and OpenShell indicates that NVIDIA’s enterprise agent strategy is not limited to general office automation.
NVIDIA’s announcement does not prove that autonomous enterprise AI agents are ready for broad unsupervised deployment. It does, however, provide a clearer view of how NVIDIA wants to package models, runtimes and developer tools for companies building them.
NVIDIA announced Nemotron 3 Ultra, a 550 billion parameter mixture of experts open model designed for long running enterprise AI agents.
The company described Nemotron 3 Ultra as an open model, positioning it for enterprises and developers that want to build AI systems with more control over deployment and integration.
The announcement fits into NVIDIA’s broader push to provide not just chips, but also model, runtime and software components for companies building AI applications.
Continue reading
