NVIDIA and LangChain Say Nemotron 3 Ultra Rivals Closed AI Models at One-Tenth the Cost

NVIDIA announced that its open Nemotron 3 Ultra model, when paired with a tuned version of LangChain’s Deep Agents framework, achieved business-task performance comparable to the highest-scoring closed models on LangChain’s Deep Agents benchmark while running at one-tenth the inference cost per run. According to the company, the results were achieved without retraining the model, with all performance gains coming from changes to the agent environment rather than the underlying model.

The collaboration centers on LangChain’s Deep Agents harness, which was specifically optimized for Nemotron 3 Ultra. LangChain said it analyzed execution traces from its public benchmark suite to identify where the model was losing points, then improved system prompts, tool descriptions and middleware instead of modifying the model itself. The tuned configuration delivered the highest accuracy among open models, completed more tasks, and operated at higher throughput than leading closed models in the benchmark.

The benchmark results suggest that engineering the broader agent system can substantially improve performance without requiring additional model training. NVIDIA said the lower operating cost allows organizations to run evaluations continuously, iterate more quickly and deploy specialized AI agents across a wider range of business workflows.

LangChain’s agent engineering platform, which the company says records more than 200 million monthly downloads, now makes the tuned Deep Agents profile for Nemotron 3 Ultra directly available to developers. The companies position the work as an example of improving the surrounding infrastructure—including memory, tool usage and evaluation—rather than relying solely on larger or retrained foundation models.

“The way to build better agents is to keep improving the system around the model,” said Harrison Chase, cofounder and CEO of LangChain. “Memory, tool use, evaluation and model behavior compound when teams can tune them together. Our work with NVIDIA shows that enterprises can get strong performance from an open stack while keeping control over the agent systems they are building.”

Alongside the benchmark results, NVIDIA introduced NemoClaw for LangChain Deep Agents, an open reference blueprint designed for enterprises building specialized AI systems. The package combines LangChain Deep Agents Code, optimized for Nemotron 3 Ultra, with NVIDIA OpenShell, a secure runtime for executing agent actions.

NVIDIA said the combination of an open model, open agent framework and open runtime gives enterprises control over the entire deployment stack, allowing organizations to customize systems for their own workflows, operate them on their preferred infrastructure and manage them under their own governance policies.

Several organizations are already adopting the approach. NVIDIA said Abridge, Amdocs and Box are embedding specialized agents into their platforms, while EY is expanding its implementation capabilities around NemoClaw blueprints to help customers customize, evaluate and govern AI agents for business workflows.

NVIDIA founder and CEO Jensen Huang also recently spoke with Chase about the recent acceleration of enterprise AI capabilities, a discussion the company highlighted alongside the announcement.

The tuned Deep Agents profile for Nemotron 3 Ultra is available now through LangChain. Developers can also build from the NemoClaw for LangChain Deep Agents blueprint or access hosted deployments of Nemotron 3 Ultra through Baseten, Crusoe Cloud, DeepInfra, Fireworks, Nebius and Together AI. EY will also offer services to help enterprises deploy specialized agents using the open software stack.

This analysis is based on reporting from NVIDIA.

Image courtesy of NVIDIA.

This article was generated with AI assistance and reviewed for accuracy and quality.