NVIDIA says tuned Nemotron 3 Ultra stack cuts agent inference cost against closed models

NVIDIA has published a new enterprise AI update around Nemotron 3 Ultra, saying a tuned LangChain Deep Agents setup can deliver stronger open-model agent performance while lowering the cost of repeated inference runs. The July 8 announcement is aimed at companies building agents that use tools, memory, evaluation loops and secure runtimes rather than simple chat interfaces.

According to NVIDIA, LangChain tuned its Deep Agents harness for Nemotron 3 Ultra and reached the highest accuracy among open models on LangChain's Deep Agents benchmark. NVIDIA also says the setup completed more tasks at higher throughput and ran at 10x lower inference cost per run than leading closed models. The company frames the result as a harness-engineering gain: LangChain analyzed execution traces and adjusted prompts, tool descriptions and middleware around the model, instead of retraining Nemotron itself.

Why It Matters

The pitch is important because enterprise AI buyers are increasingly weighing performance against control. Closed frontier models can be convenient, but they may limit how much a company can customize the surrounding system, inspect behavior, govern data flow or run workloads on its own infrastructure. NVIDIA is presenting Nemotron 3 Ultra, LangChain Deep Agents and its OpenShell secure runtime as an open stack that businesses can tune, operate and govern more directly.

NVIDIA says the tuned profile is available now through LangChain, while the NemoClaw for LangChain Deep Agents blueprint is available as a starting point for specialized agents. The company lists hosted access paths for Nemotron 3 Ultra through Baseten, Crusoe Cloud, DeepInfra, Fireworks, Nebius and Together AI. It also names Abridge, Amdocs, Box and EY as organizations embedding or supporting specialized agents around the stack.

• The main claim is not that a new model was trained, but that orchestration changes improved agent results.
• The focus is business agents that execute tasks across tools and systems, where cost can compound quickly.
• The announcement keeps pressure on closed-model providers by emphasizing open deployment and governance options.

As with most vendor benchmarks, the reported performance and cost comparisons should be read as NVIDIA's characterization of a specific benchmark environment. Still, the release signals where enterprise AI competition is moving: from raw model releases toward complete agent systems that can be measured, secured and adapted for production work.

Sources

• blogs.nvidia.com

Cover photo by Sergei Starostin on Pexels, used under the Pexels License.