New NVIDIA Nemotron 3 Super Delivers 5x Higher Throughput for Agentic AI
Key Facts
- What: NVIDIA launched Nemotron 3 Super, a 120-billion-parameter open hybrid mixture-of-experts model with 12 billion active parameters optimized for agentic AI workflows.
- When: Launched March 11, 2026.
- Performance: Delivers up to 5x higher throughput and up to 2x higher accuracy than the previous Nemotron Super model, with a 1-million-token context window.
- Architecture: Combines Mamba layers for efficiency, transformer layers for reasoning, latent MoE, and multi-token prediction.
- Availability: Open weights released under a permissive license; optimized to run on NVIDIA Blackwell in NVFP4 precision.
NVIDIA on March 11, 2026, unveiled Nemotron 3 Super, a new open 120-billion-parameter hybrid mixture-of-experts model designed to overcome the high costs and performance bottlenecks of complex autonomous agent systems. The model addresses two major constraints in multi-agent AI workflows — context explosion and the “thinking tax” — by delivering up to 5x higher throughput while maintaining leading accuracy.
Perplexity is immediately offering users access to Nemotron 3 Super for search and as one of 20 orchestrated models in its Computer product. Multiple software development agents, including those from CodeRabbit, Factory and Greptile, are integrating the model alongside proprietary systems to improve accuracy at lower cost. Life sciences organizations Edison Scientific and Lila Sciences also plan to use it for deep literature search, data science and molecular understanding.
Addressing Agentic AI Challenges
As enterprises move from simple chatbots to sophisticated multi-agent applications, they face significant scaling issues. Multi-agent workflows can generate up to 15x more tokens than standard chat interactions because each step requires resending full conversation histories, tool outputs and intermediate reasoning. This “context explosion” drives up costs and increases the risk of goal drift, where agents lose alignment with their original objectives.
The second major constraint is the “thinking tax.” Complex agents must perform advanced reasoning at every step, but routing every subtask through large, dense models makes these systems too expensive and slow for real-world deployment.
Nemotron 3 Super tackles both problems with a 1-million-token context window that lets agents maintain complete workflow state in memory, reducing the need to repeatedly reprocess history. According to NVIDIA’s announcement, the model has claimed the top spot on Artificial Analysis for efficiency and openness while delivering leading accuracy among models of comparable size. It also powers the NVIDIA AI-Q research agent to the No. 1 position on both DeepResearch Bench and DeepResearch Bench II leaderboards, which evaluate multistep research capabilities across large document collections.
Hybrid Architecture and Efficiency Gains
The new model employs a hybrid mixture-of-experts (MoE) architecture that integrates several technical innovations. Mamba layers provide 4x higher memory and compute efficiency for handling long contexts, while transformer layers preserve strong reasoning performance. Only 12 billion of the model’s 120 billion parameters are active during inference.
A new “latent MoE” technique activates four expert specialists for the computational cost of one when generating each token, improving accuracy without proportional increases in compute. Multi-token prediction allows the model to forecast multiple future tokens simultaneously, resulting in 3x faster inference.
When running on the NVIDIA Blackwell platform in NVFP4 precision, the model reduces memory requirements and achieves up to 4x faster inference compared with FP8 precision on NVIDIA Hopper, according to NVIDIA, with no loss in accuracy.
Open Release and Training Transparency
NVIDIA is releasing Nemotron 3 Super with open weights under a permissive license, enabling developers to deploy and customize the model on workstations, in data centers or in the cloud. The company trained the model using synthetic data generated by frontier reasoning models and is publishing the full methodology, including over 10 trillion tokens of pre- and post-training datasets, 15 training environments for reinforcement learning, and complete evaluation recipes.
Developers can further fine-tune the model using the NVIDIA NeMo platform. The release is part of NVIDIA’s broader effort to support the growing ecosystem of agentic AI applications across industries.
Industry Adoption and Use Cases
Enterprise software platforms are already deploying the model. Companies including Amdocs, Palantir, Cadence, Dassault Systèmes and Siemens are customizing Nemotron 3 Super to automate workflows in telecom, cybersecurity, semiconductor design and manufacturing.
In software development, the model’s long context window allows agents to load entire codebases at once for end-to-end code generation and debugging without segmentation. In financial analysis, it can process thousands of pages of reports in a single context, eliminating redundant reasoning across long conversations. The model’s high-accuracy tool-calling capabilities are particularly valuable in high-stakes environments such as autonomous security orchestration in cybersecurity.
Impact
For developers and enterprises building agentic systems, Nemotron 3 Super offers a path to significantly lower inference costs while maintaining or improving accuracy. The combination of massive context length and high throughput makes previously impractical multi-agent workflows economically viable. Early adopters in software development, life sciences and enterprise automation are expected to see immediate benefits in both performance and operating expenses.
The open nature of the release, including full training data and methodology, should accelerate research and customization across the AI community. By publishing complete recipes, NVIDIA aims to enable researchers and organizations to build upon the work rather than starting from scratch.
What's Next
NVIDIA has indicated that additional models in the Nemotron 3 family are planned. The current Super release focuses on high-throughput agentic workloads, while future variants may emphasize different trade-offs between accuracy, speed and scale.
The model is available immediately for download and integration. Organizations can begin testing it on NVIDIA Blackwell systems or compatible infrastructure to evaluate its performance in their specific agentic AI applications.
Sources
- New NVIDIA Nemotron 3 Super Delivers 5x Higher Throughput for Agentic AI | NVIDIA Blog
- NVIDIA Debuts Nemotron 3 Family of Open Models | NVIDIA Newsroom
- Nemotron AI Models | NVIDIA Developer
- Inside NVIDIA Nemotron 3: Techniques, Tools, and Data That Make It Efficient and Accurate | NVIDIA Technical Blog
- NVIDIA Nemotron 3 Family of Models - NVIDIA Nemotron
All technical specifications, pricing, and benchmark data in this article are sourced directly from official announcements. Competitor comparisons use publicly available data at time of publication. We update our coverage as new information becomes available.