NVIDIA Nemotron 3 Ultra Tops Open Model Accuracy With LangChain Harness

NVIDIA announced that its Nemotron 3 Ultra model, when paired with a LangChain-tuned Deep Agents harness, achieves the highest accuracy among open models on the Deep Agents benchmark. The company also claims the model completes more tasks, delivers higher throughput, and cuts inference cost per run by 10x compared with leading closed models. NVIDIA says Nemotron 3 Ultra matches the business-task performance of the highest-scoring closed models on the same benchmark.

Engineering the environment, not the model

The performance gains came without any model retraining. LangChain's team analyzed execution traces from agent runs and tuned prompts, tool descriptions, and middleware instead of modifying the model weights. Harrison Chase, LangChain's co-founder, said improving the system around the model improves agent performance. NVIDIA frames this as a shift from model-centric optimization to environment-level engineering, which lets teams run continuous evaluations at one-tenth the cost using Nemotron 3 Ultra.

The tuned profile is available directly through LangChain for any developer. LangChain's agent engineering platform records over 200 million monthly downloads, according to the company. Developers can pull the tuned Deep Agents harness from LangChain or use the NemoClaw blueprint to build agents from scratch.

NemoClaw: an open reference blueprint

NVIDIA also introduced NemoClaw for LangChain Deep Agents, an open reference blueprint for building specialized AI stacks. NemoClaw combines the LangChain Deep Agents code tuned for Nemotron 3 Ultra with the OpenShell secure runtime. The open model, harness, and runtime let enterprises own, customize, and run the full stack anywhere. NVIDIA says the tuned harness enables high-performing agents that complete more tasks, run faster, and provide a fully open stack.

NemoClaw and the tuned Nemotron 3 Ultra model profile are available now. Nemotron 3 Ultra is accessible on Baseten, Crusoe Cloud, DeepInfra, Fireworks, Nebius, and Together AI platforms. EY can help enterprises start building specialized agents using the open software stack. Companies like Abridge, Amdocs, and Box embed specialized agents, and EY expands its NVIDIA implementation with NemoClaw blueprints. Jensen Huang discussed the recent enterprise AI leap with Harrison Chase.

For inference engineers, the key takeaway is that agent performance can be improved by tuning the harness, not just the model. The 10x cost reduction per run, if it holds in practice, makes continuous evaluation more feasible for teams serving open-weight models. The open stack also removes dependency on closed APIs, which may matter for enterprises that need to run agents on their own infrastructure.