Nvidia’s Nemotron 3 Super: A Leap Forward for Agentic AI
Nvidia has introduced Nemotron 3 Super, a 120-billion-parameter hybrid model designed to address the computational demands of multi-agent systems and long-horizon tasks like software engineering and cybersecurity. The model’s weights are now available on Hugging Face, offering a commercially usable solution with a focus on efficiency and transparency.
Addressing the Context Explosion
Multi-agent systems, capable of handling complex tasks, often generate significantly more tokens than standard chat applications—potentially increasing costs. Nemotron 3 Super aims to mitigate this issue by providing specialized depth for agentic workflows without the computational overhead typically associated with dense reasoning models.
A Triple Hybrid Architecture
At the heart of Nemotron 3 Super lies a sophisticated architecture that balances memory efficiency with precise reasoning. This architecture combines three key elements:
- Hybrid Mamba-Transformer Backbone: This backbone interleaves Mamba-2 layers with Transformer attention layers. Mamba-2 layers enable fast processing of sequences with linear-time complexity, supporting a massive 1-million-token context window. Transformer attention layers act as “global anchors,” ensuring accurate retrieval of specific information.
- Latent Mixture-of-Experts (LatentMoE): Traditional Mixture-of-Experts (MoE) designs can create computational bottlenecks as models scale. LatentMoE addresses this by projecting tokens into a compressed space before routing them to specialized experts, allowing for more efficient use of computational resources.
- Multi-Token Prediction (MTP): Unlike standard models that predict one token at a time, MTP predicts multiple future tokens simultaneously. This “built-in draft model” enables native speculative decoding, potentially delivering up to 3x faster performance for structured generation tasks like code or tool calls.
Blackwell Optimization and Performance Gains
Nemotron 3 Super is optimized for the Nvidia Blackwell GPU platform. By pre-training natively in NVFP4 (4-bit floating point), Nvidia has achieved significant production efficiency gains. On Blackwell, the model delivers 4x faster inference than 8-bit models running on the previous Hopper architecture, with no loss in accuracy.
In benchmark testing, Nemotron 3 Super currently ranks No. 1 on the DeepResearch Bench, which measures an AI’s ability to conduct thorough, multi-step research across large document sets. It also demonstrates significant throughput advantages, achieving up to 2.2x higher throughput than gpt-oss-120B and 7.5x higher than Qwen3.5-122B in high-volume settings.
Licensing and Ecosystem Adoption
The model is released under the Nvidia Open Model License Agreement (updated October 2025), which allows for commercial usage with specific safeguard clauses. Key provisions include commercial usability, ownership of output, and the freedom to create derivative works, provided attribution is included.
The license includes termination triggers related to safety guardrails and potential litigation. This structure aims to foster a commercial ecosystem while protecting Nvidia’s intellectual property and ensuring responsible use of the model.
Industry Integration
Nemotron 3 Super is being deployed as an Nvidia NIM microservice, enabling it to run on-premises via Dell AI Factory or HPE, as well as across Google Cloud, Oracle, and soon, AWS and Azure. Companies like CodeRabbit and Greptile are integrating the model for large-scale codebase analysis, while Siemens and Palantir are deploying it to automate complex workflows in manufacturing and cybersecurity.
As Nvidia VP of AI Software, Kari Briski, noted, Nemotron 3 Super addresses the “context explosion” encountered by companies moving beyond chatbots into multi-agent applications, providing the necessary “brainpower” with improved operational efficiency.