Nvidia's Nemotron 3 Ultra tops open-source models in Deep Agents benchmark

Nvidia announced that its Nemotron 3 Ultra model achieved the highest accuracy among open-source models in LangChain's Deep Agents benchmark. Deep Agents is a system that handles multi-step reasoning, uses external tools, and completes complex tasks, often used for workflow automation and cross-system collaboration. Nvidia noted that the test focused on cost and throughput. The optimized Nemotron 3 Ultra runs at one-tenth the cost of leading closed-source models per operation, while completing more tasks with higher throughput. In business task capabilities, the model matched top closed-source models by optimizing its surrounding systems. LangChain CEO Harrison Chase said building better agents depends on improving systems around the model, like memory and tool use. The partnership shows companies can get strong performance with open tech stacks while controlling their agent systems.