NVIDIA says Nemotron 3 Ultra, tuned with LangChain Deep Agents, hits benchmark-level performance at lower cost

NVIDIA is stepping up its push for practical “agentic” AI with a new set of claims that its Nemotron 3 Ultra model can match top closed rivals on a widely used evaluation for deep agents, while lowering inference costs enough to support continuous testing and faster iteration. In a July 8 post, the company says LangChain tuned its Deep Agents harness specifically for Nemotron 3 Ultra and found the resulting setup achieved the highest accuracy among open-model options tested on LangChain’s Deep Agents benchmark suite.

The companies frame the work as an emphasis on system design around a model, rather than altering the model itself. NVIDIA says the benchmark gains came from engineering the environment around the model, including adjustments to system prompts, tool descriptions, and middleware, after tracing where the agent execution lost points during benchmark runs. It adds that no model retraining was required for the tuned performance to emerge.

NVIDIA ties the benchmark result to task outcomes by saying Nemotron 3 Ultra achieved “business task parity” with the highest-scoring closed models in the same Deep Agents benchmark, while also completing more tasks at higher throughput. The post does not publish the detailed benchmark scoreboard in its text, but it does describe the evaluation as one that measures agent execution quality across tasks rather than only raw language quality.

Cost is a central theme in the release. NVIDIA says LangChain’s tuning lets teams run at “10x lower inference cost per run” than leading closed models, and it argues that at a tenth of the cost, developers can keep running evaluations continuously, experiment more quickly, and build specialized agents for more workflows. NVIDIA also characterizes the approach as reducing friction for teams trying to move from prototype agent assistants toward agents that operate inside enterprise systems.

NVIDIA further says the LangChain Deep Agents ecosystem is already large, pointing to LangChain’s agent engineering platform having more than 200 million monthly downloads. By tuning the Deep Agents harness for Nemotron 3 Ultra, the companies say enterprises can assemble high-performing agents using an open stack that they can customize, own, and run across their own infrastructure or preferred clouds.

The release also spotlights an open blueprint intended to package the pieces for enterprise deployments. NVIDIA describes “NemoClaw for LangChain Deep Agents” as an open reference configuration that combines LangChain Deep Agents Code, tuned for Nemotron 3 Ultra, with NVIDIA OpenShell, a secure runtime for executing agent actions safely. In this framing, the “stack” is not only the model, but also the tools, runtime, and runtime controls that govern what an agent is allowed to do when it takes action.

Enterprise adoption indicates appear in the post through specific partners and customers embedding agent capabilities. NVIDIA names Abridge, Amdocs, and Box as companies embedding specialized agents into their platforms and global systems, and it says EY is expanding its NVIDIA implementation capabilities around NVIDIA NemoClaw blueprints for LangChain Deep Agents. NVIDIA characterizes that work as helping clients customize, evaluate, and govern specialized agents across high-value workflows.

NVIDIA’s release also provides a mechanism for how developers can start using the tuned setup. It says the tuned harness profile for Nemotron 3 Ultra is available directly through LangChain, and that developers can access Nemotron 3 Ultra through multiple hosted platforms including Baseten, Crusoe Cloud, DeepInfra, Fireworks, Nebius, and Together AI, described as providing a direct path to the tuned harness in production.

One caveat in the release is that it does not include the underlying benchmark methodology, pass rates, or the specific closed-model baselines used in the “10x lower cost” and “business task parity” claims. It also does not specify absolute dollar amounts, the hardware or runtime configuration used for inference cost comparisons, or the scope of the benchmark tasks included in the “Deep Agents benchmark suite.” As a result, readers are left with qualitative claims about performance and cost, rather than a fully auditable set of numbers in the post itself.

Looking ahead, the key question for enterprises is whether this “tune the harness, not the model” approach generalizes across different tool ecosystems and enterprise workflows. NVIDIA and LangChain say NemoClaw and the tuned Nemotron 3 Ultra profile are available now, so teams evaluating agentic AI may focus on testing continuous evaluation loops, governance controls, and cost profiles under their own constraints. Monitoring how quickly developers adopt the open blueprint and how partners operationalize “deep agents” in production workflows could indicate whether the benchmark-style advantages translate into sustained enterprise deployments.