NVIDIA has introduced NVIDIA Nemotron 3 Super, a 120-billion-parameter open model designed to support large-scale agentic AI systems capable of executing complex, multi-step tasks with improved efficiency and accuracy.

The newly launched model features 12 billion active parameters and a 1-million-token context window, enabling AI agents to maintain extended workflow memory while handling large volumes of information. The architecture is designed to address key challenges in emerging multi-agent AI applications, including rising computational costs and context management.

As enterprises move beyond chatbot-style interactions toward autonomous multi-agent applications, AI systems must manage significantly larger context volumes. Multi-agent workflows can generate up to 15 times more tokens than traditional chat interactions because each step requires full histories of tool outputs, intermediate reasoning and prior prompts.

This “context explosion” can increase costs and cause goal drift, where agents lose alignment with the original task. According to NVIDIA, Nemotron 3 Super’s extended context window allows agents to retain the full workflow state, improving reasoning consistency across long-running tasks.

The model is also designed to address the so-called “thinking tax”—the high computational cost of applying large models to every subtask within complex workflows. By balancing reasoning capability with efficiency, the system aims to enable practical deployment of scalable multi-agent applications.

Several technology companies have already begun integrating the model into their platforms. AI-native search company Perplexity AI is providing access to Nemotron 3 Super for search and as one of the orchestrated models powering its AI system “Computer.”

Meanwhile, developer-focused platforms including CodeRabbit, Factory AI and Greptile are incorporating the model into software development agents alongside proprietary models to improve accuracy and cost efficiency.

The model is also being adopted in scientific and research applications. Organizations such as Edison Scientific and Lila Sciences plan to deploy the system to power AI agents capable of deep literature analysis, data science workflows and molecular research.

Major enterprise software providers including Amdocs, Palantir Technologies, Cadence Design Systems, Dassault Systèmes and Siemens are also deploying and customizing the model for applications ranging from telecom automation and cybersecurity to semiconductor design and industrial manufacturing workflows.

NVIDIA said Nemotron 3 Super has achieved leading efficiency and openness scores on the Artificial Analysis benchmark, outperforming models of similar size. The model also powers the NVIDIA AI‑Q research agent, which has reached the top position on the DeepResearch Bench and DeepResearch Bench II leaderboards—benchmarks that evaluate AI systems’ ability to conduct complex, multi-step research across large document collections while maintaining coherent reasoning.

The launch highlights NVIDIA’s growing focus on agentic AI, a new generation of systems designed to autonomously plan, reason and execute tasks across multiple tools and data sources.