NVIDIA Nemotron 3 Ultra sets an accuracy record among open models in LangChain test

NVIDIA and LangChain presented in July 2026 the results of joint tuning of the agent framework: Nemotron 3 Ultra demonstrated the highest accuracy among all open-source models in LangChain's Deep Agents test — while outperforming leading closed models in the price-to-quality ratio.

What is LangChain Deep Agents?

LangChain is the world's largest and most widely used platform for orchestrating AI agents. Millions of developers build agent applications on its foundation: from business process automation to complex multi-step data analysis.

Deep Agents is a specialized LangChain framework for agents with deep planning. Unlike simple prompt chains, Deep Agents allows the model to independently break down a task into subtasks, select tools, iterate on intermediate results, and correct the plan as execution proceeds. This mode is one of the most demanding on language model quality: each intermediate error accumulates and reduces overall accuracy.

NVIDIA Nemotron 3 Ultra is an open-source language model from the Nemotron family, oriented toward enterprise deployment. The LangChain team not only integrated it into its stack but specially tuned the framework to the architecture of this model — which enabled achieving record results in agent mode.

What Nemotron 3 Ultra Achieved in the Benchmark

The tuning results proved significant for the entire open-source model market:

• Accuracy: highest among all open-source models in the Deep Agents harness test
• Cost: lower than leading closed competitors with comparable quality
• Throughput: more tasks completed per unit time
• Magnitude of advantage: the model outperforms alternatives by 10x on one of the key operational metrics

Historically, closed models — GPT-4o, Claude, Gemini — led in complex agent tasks over open-source ones. An agent performs dozens of iterations, calls external tools, and maintains context between steps: the cost of each intermediate failure is high. Nemotron 3 Ultra coupled with the Deep Agents harness tuned for it breaks this barrier for open-source models for the first time.

"NVIDIA

Nemotron 3 Ultra offers leading performance at lower cost than top closed models."

Why is an Open Model Better than a Closed API?

An open model with the performance level of closed APIs provides corporate developers with a number of fundamental advantages. First and foremost — data does not leave the company's infrastructure. Banks, government sector, and healthcare face strict data localization requirements — this is precisely where open models gain the greatest practical advantage over API services.

Moreover, an open model eliminates dependence on a third-party provider's pricing policy: there is no risk of sudden rate increases or changes in model behavior. Inference costs can be optimized for actual workloads, and if needed, the model can be fine-tuned on corporate data.

Open models are also more cost-effective when running agents under high load: deployment on your own cluster allows linear scaling of capacity without increasing API call costs.

Not insignificant is the ecosystem effect: LangChain is used by millions of developers, which means the entire layer of agent applications can now run on an open model without loss of performance.

What This Means

The victory of an open model in the benchmark of the world's largest agent platform shifts the balance of power: closed APIs cease to be a necessary condition for high-performance agent applications. For developers already using LangChain, Nemotron 3 Ultra becomes a real alternative — at lower cost and without sacrificing accuracy.