NVIDIA Nemotron Achieves Benchmark-Leading Performance With LangChain Deep Agents Harness
NVIDIA Blog · Adel El Hallak · 2026-07-08
NVIDIA and LangChain announce that Nemotron 3 Ultra achieves top accuracy among open models on the Deep Agents benchmark at 10x lower inference cost than leading closed models, with all gains attributed to harness engineering rather than model retraining.
Appears in
Extraction
Topics: ai-agentsopen-source-modelsinference-costai-benchmarksenterprise-ai
Claims
- LangChain's Deep Agents harness tuned for Nemotron 3 Ultra achieves the highest accuracy among open models and reaches business task parity with top closed models on the Deep Agents benchmark.
- Nemotron 3 Ultra runs at 10x lower inference cost per run than leading closed models while matching their task completion performance.
- All performance gains came from engineering the harness environment—adjusting system prompts, tool descriptions, and middleware—with no model retraining required.
- LangChain's agent engineering platform has more than 200 million monthly downloads.
- The NVIDIA NemoClaw blueprint provides an open reference stack combining an open model, open harness, and secure open runtime that enterprises can own and run anywhere.
Key quotes
The way to build better agents is to keep improving the system around the model.
Every gain came from engineering the environment around the model, not the model itself.
At a tenth of the cost, teams harnessing NVIDIA Nemotron 3 Ultra can run evaluations continuously, experiment faster and build specialized agents across more of their business.