Managing the economics of multi-agent AI now dictates the financial viability of modern business automation workflows.

Organisations progressing past standard chat interfaces into multi-agent applications face two primary constraints: the thinking tax and context explosion. Complex autonomous agents need to reason at each stage, making the reliance on massive architectures too expensive and slow. These advanced workflows produce up to 1,500 percent more tokens than standard formats, driving up expenses and causing goal drift.

Architectures for Multi-Agent AI

To address these hurdles, Nvidia released Nemotron 3 Super, an open architecture featuring 120 billion parameters (of which 12 billion remain active), engineered for complex agentic AI systems. This framework blends advanced reasoning features to improve efficiency and accuracy in business automation.

The system relies on a hybrid mixture-of-experts architecture, with Mamba layers for four times the memory and compute efficiency, and standard transformer layers to manage complex reasoning requirements. A latent technique boosts accuracy, while the system anticipates multiple future words, accelerating inference speeds threefold.

Operating on the Blackwell platform, the architecture utilises NVFP4 precision, reducing memory needs and making inference up to four times faster than FP8 configurations on Hopper systems.

Automation and Business Outcomes

The system offers a one-million-token context window, allowing agents to keep the entire workflow state in memory. A software development agent can load an entire codebase into context, enabling end-to-end code generation and debugging without segmentation. Within financial analysis, the system can load thousands of pages of reports, improving efficiency. High-accuracy tool calling ensures autonomous agents reliably navigate massive function libraries, preventing errors in critical environments.

Industry leaders – including Amdocs, Palantir, Cadence, Dassault Systèmes, and Siemens – are deploying and customising the model to automate workflows across telecom, cybersecurity, semiconductor design, and manufacturing. Software development platforms like CodeRabbit, Factory, and Greptile are integrating it to achieve higher accuracy at lower costs. Life sciences firms like Edison Scientific and Lila Sciences will use it for deep literature search, data science, and molecular understanding.

The architecture claimed the top position on DeepResearch Bench and DeepResearch Bench II leaderboards, highlighting its capacity for multistep research across large document sets. It also claimed the top spot on Artificial Analysis for efficiency and openness.

Implementation and Infrastructure Alignment

Nvidia released the model with open weights under a permissive license, letting developers deploy and customise it across workstations, data centres, or cloud environments. It is packaged as an NVIDIA NIM microservice to aid deployment from on-premises systems to the cloud. For those evaluating on-premise deployments, there are trade-offs to consider. AI-RADAR offers analytical frameworks on /llm-onpremise to evaluate these aspects.

The architecture was trained on synthetic data generated by frontier reasoning models. Nvidia published the complete methodology, encompassing over 10 trillion tokens of pre- and post-training datasets, 15 training environments for reinforcement learning, and evaluation recipes. Researchers can further fine-tune the model or build their own using the NeMo platform.