NVIDIA Launches Nemotron 3 Super, 120B Open Model Optimized for Agentic AI
Key Facts
- What: NVIDIA released Nemotron 3 Super, a 120-billion-parameter open hybrid SSM Latent MoE model with 12 billion active parameters
- Architecture: Hybrid Mamba-Transformer Mixture-of-Experts designed specifically for Blackwell GPUs
- Performance: Delivers up to 5x higher throughput for agentic AI workloads; achieves 36 on AAIndex v4
- Context Length: Supports up to 1 million tokens
- Availability: Open model weights released; available immediately through multiple inference providers including Baseten, Cloudflare, DeepInfra, Fireworks AI and others
- Use Case: Optimized for complex, collaborative multi-agent AI systems at scale
NVIDIA has unveiled Nemotron 3 Super, a powerful new open-source AI model engineered to run complex agentic AI systems at unprecedented scale and efficiency. The 120-billion-parameter hybrid model, which activates only 12 billion parameters per token, is designed specifically for Blackwell architecture and promises up to 5x higher throughput compared to previous approaches for multi-agent workloads.
Announced today via NVIDIA's official AI developer channels, Nemotron 3 Super represents the flagship offering in the expanded Nemotron 3 family. It combines state-space model (SSM) elements with transformer and Mixture-of-Experts (MoE) architecture to deliver both high reasoning accuracy and exceptional inference efficiency — critical requirements for building collaborative AI agents that can handle sophisticated, long-context tasks.
Technical Architecture and Design
According to NVIDIA's technical blog, Nemotron 3 Super employs a novel Hybrid SSM Latent MoE architecture. This 120B-12A configuration (120 billion total parameters with 12 billion active) integrates Mamba-style state-space layers with traditional transformer attention mechanisms and a Mixture-of-Experts routing system.
The hybrid design allows the model to leverage the linear scaling advantages of Mamba for long sequences while retaining the strong reasoning capabilities of transformers. NVIDIA specifically optimized the model for its Blackwell GPU architecture, enabling significant performance gains on next-generation hardware.
Key technical specifications include:
- Total Parameters: 120 billion
- Active Parameters per Token: 12 billion
- Maximum Context Length: 1 million tokens
- Architecture: Hybrid Mamba-Transformer MoE (SSM Latent MoE)
- Benchmark Score: 36 on AAIndex v4
The model is positioned as a high-accuracy reasoning system particularly suited for multi-agent applications. Unlike dense models that activate all parameters for every token, the MoE design dramatically reduces computational requirements during inference while maintaining strong performance on complex reasoning tasks.
Performance and Benchmarks
NVIDIA claims Nemotron 3 Super delivers up to 5x higher throughput for agentic AI workloads compared to traditional dense models of similar capability. This efficiency gain is especially important for multi-agent systems, where multiple AI instances must collaborate in real time on complex tasks.
The model achieved a score of 36 on AAIndex v4, a benchmark that evaluates agentic AI capabilities. While specific comparisons to other leading models were not detailed in the initial announcement, the score positions Nemotron 3 Super as a competitive entrant in the emerging category of specialized agentic foundation models.
Its 1-million-token context window is particularly notable, allowing the model to maintain coherence across extremely long conversations, codebases, or document collections — a crucial capability for autonomous agent workflows that may need to reference extensive historical context or documentation.
Availability and Ecosystem Support
NVIDIA has released Nemotron 3 Super as an open model, making the weights available to developers and researchers. The company has partnered with numerous inference providers to ensure immediate accessibility:
- Baseten
- Cloudflare
- DeepInfra
- Fireworks AI
- FriendliAI
- Inference.net
- Lightning AI
- Modal
- Nebius
- Together AI
Detailed getting-started instructions are available in NVIDIA's GitHub repository, including integration guides for popular agentic frameworks such as OpenCode, OpenHands, and OpenClaw. A comprehensive technical report has also been published for researchers seeking deeper insights into the model's architecture and training methodology.
The release builds upon NVIDIA's earlier introduction of the broader Nemotron 3 family, which includes smaller models like Nemotron 3 Nano. The "Super" tier specifically targets high-volume, collaborative agent workloads — a positioning that remains consistent with NVIDIA's description of the model family when it was first unveiled last year.
Competitive Context in the Agentic AI Landscape
Nemotron 3 Super enters a rapidly evolving market for foundation models optimized for agentic workflows. Major players including OpenAI, Anthropic, Google, and Meta have all been investing heavily in models capable of powering autonomous AI agents that can reason, plan, and collaborate to accomplish complex goals.
NVIDIA's approach differs by focusing on inference efficiency and massive context windows through its hybrid architecture. By leveraging both Mamba and transformer components in an MoE framework, the company aims to deliver strong reasoning performance at significantly lower operational cost than purely transformer-based alternatives.
This focus on efficiency is particularly strategic given the explosive computational demands of multi-agent systems. As organizations begin deploying fleets of specialized AI agents working together, the ability to run these systems at scale without prohibitive infrastructure costs becomes a major competitive advantage.
The model's optimization for Blackwell GPUs also aligns with NVIDIA's hardware roadmap, potentially giving customers who invest in the company's latest accelerators a performance edge when deploying agentic AI applications.
Impact on Developers and Enterprise AI
For developers building AI applications, Nemotron 3 Super offers a powerful new open option for agentic systems. The combination of high reasoning accuracy, massive context length, and strong throughput makes it suitable for sophisticated use cases including:
- Autonomous software development agents
- Complex research and data analysis workflows
- Multi-step business process automation
- Collaborative AI systems that coordinate multiple specialized agents
The open nature of the model allows organizations to fine-tune and deploy it within their own infrastructure, addressing data privacy and customization requirements that might prevent adoption of closed-source alternatives.
Enterprise users particularly stand to benefit from the efficiency improvements. The 5x throughput gain for agentic workloads could translate into substantially lower inference costs and higher scalability for production deployments involving multiple collaborating agents.
What's Next for Nemotron 3 and Agentic AI
NVIDIA has not announced specific timelines for future updates to the Nemotron 3 family, though the rapid pace of advancement in agentic AI suggests additional model variants and capability improvements are likely in development.
The company has emphasized the importance of the broader ecosystem, with GitHub resources and partnerships with inference providers designed to lower barriers to adoption. As more developers experiment with Nemotron 3 Super in frameworks like OpenHands and OpenClaw, the community may develop best practices and additional tooling that further accelerates agentic AI development.
The release also reinforces NVIDIA's growing role not just as a hardware provider but as a key player in the AI software and model ecosystem. By releasing high-performance open models optimized for its hardware, NVIDIA strengthens the value proposition of its GPU platforms for the next generation of AI workloads.
Industry observers will be watching closely to see how Nemotron 3 Super performs against competing agentic models in independent benchmarks and real-world deployments. Its hybrid architecture represents an interesting technical bet on the future of efficient, long-context reasoning systems.
As agentic AI moves from research curiosity to production reality, models like Nemotron 3 Super that specifically target the unique computational patterns of multi-agent collaboration could prove instrumental in making sophisticated autonomous systems practical at scale.
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
- Introducing Nemotron 3 Super: An Open Hybrid Mamba-Transformer MoE for Agentic Reasoning | NVIDIA Technical Blog
- Nvidia releases Nemotron 3 Super, 120B open model with 1M-token context - VideoCardz.com
- Nvidia launches Nemotron 3 Super AI model for agentic AI systems | Seeking Alpha
- Original announcement via @NVIDIAAIDev on X