NVIDIA Releases Nemotron 3 Super: Open 120B Hybrid Model Targets Agentic AI
Key Facts
- What: NVIDIA launched Nemotron 3 Super, a 120-billion-parameter open hybrid Mamba-Transformer Mixture-of-Experts model with 12 billion active parameters.
- When: Launched today and immediately available on build.nvidia.com, Hugging Face, Perplexity, and OpenRouter.
- Focus: Designed for complex agentic AI systems, delivering up to 5x higher throughput compared to dense transformer models of similar capability.
- Release Scope: NVIDIA is open-sourcing the complete training and evaluation recipe from pretraining through alignment.
- Architecture: Hybrid Mamba-Transformer MoE enabling efficient scaling for reasoning and agentic workloads.
NVIDIA has unveiled Nemotron 3 Super, a 120-billion-parameter open-source AI model specifically engineered for large-scale agentic reasoning systems. The new model combines Mamba, Transformer, and Mixture-of-Experts (MoE) architectures in a hybrid design that activates only 12 billion parameters during inference, significantly improving efficiency for complex AI agent deployments.
According to NVIDIA’s official announcements, the model is now available for download and testing across major platforms including Hugging Face, Perplexity, OpenRouter, and NVIDIA’s own build.nvidia.com portal. The company is also releasing the full training pipeline documentation, allowing developers to reproduce results, create domain-specific variants, or experiment with similar hybrid architectures.
Technical Architecture and Design Choices
Nemotron 3 Super represents NVIDIA’s continued investment in hybrid architectures that move beyond traditional dense transformer models. The model leverages a Mamba-Transformer MoE design, which according to NVIDIA’s technical blog combines the linear-time sequence modeling strengths of Mamba with the strong reasoning capabilities of transformers, while using Mixture-of-Experts to reduce computational overhead.
This hybrid approach is particularly well-suited for agentic AI workloads — systems where AI agents must maintain long context, perform multi-step reasoning, and interact dynamically with tools and environments. NVIDIA reports the model achieves up to 5x higher throughput than comparable dense models when running these complex agentic tasks, addressing one of the primary bottlenecks in deploying sophisticated AI agents at scale.
The release includes not just model weights but the entire training and evaluation recipe covering pretraining, supervised fine-tuning, and alignment stages. This level of transparency is significant in the open-source AI landscape, as it enables researchers to fully understand and build upon NVIDIA’s methodology rather than treating the model as a black box.
Community and Early Testing Feedback
Early testing of the model has generated positive reactions within the AI developer community. In a post that was amplified by NVIDIA’s AI developer account, tester Sam Hogan stated the model has been under evaluation for several weeks and described it as “easily the best Open Source American model for its” size class, though the full quote was truncated in the shared update.
This feedback aligns with NVIDIA’s positioning of Nemotron 3 Super as a leading open alternative for developers seeking high-performance models without relying exclusively on closed-source offerings from companies like OpenAI, Anthropic, or Google. The “American model” reference appears to highlight its development within the U.S. technology ecosystem amid growing interest in domestic AI supply chains.
Availability and Developer Access
Developers can immediately access Nemotron 3 Super through multiple channels:
- NVIDIA’s build.nvidia.com platform
- Hugging Face model hub
- Perplexity’s platform
- OpenRouter’s routing service
The broad availability strategy reflects NVIDIA’s goal of accelerating adoption of the model across different developer workflows and inference providers. By making the model available on established platforms, NVIDIA reduces friction for teams already working within familiar ecosystems.
The complete training recipe release further lowers barriers for organizations wanting to customize the model. Companies and research groups can adapt the pipeline for specialized domains such as scientific research, financial analysis, software development, or enterprise automation where agentic capabilities provide particular value.
Industry Context and Competitive Landscape
NVIDIA’s release of Nemotron 3 Super comes at a time of intense competition in the open-source large language model space. Meta has led with its Llama series, while organizations like Mistral, Alibaba, and Snowflake have also released competitive open models. NVIDIA’s entry brings its signature focus on inference efficiency and hardware optimization, leveraging the company’s deep expertise in GPU acceleration.
The emphasis on agentic AI reflects a broader industry shift from simple chat-based applications toward more autonomous AI systems capable of planning, tool use, and multi-step problem solving. As enterprises look to deploy AI agents for complex workflows, models that can efficiently handle long contexts and sophisticated reasoning while maintaining reasonable computational costs become increasingly valuable.
NVIDIA’s hybrid Mamba-Transformer approach also signals growing interest in architectures beyond pure transformers. Models incorporating Mamba’s state-space capabilities have shown promise in efficiently handling very long sequences, making them attractive for agentic applications that may require maintaining extensive conversation history or working memory.
Impact on Developers and Enterprise AI
For developers, Nemotron 3 Super offers a powerful new option for building production-grade agentic systems without incurring the costs or dependency risks associated with proprietary APIs. The model’s efficiency characteristics — particularly the 5x throughput improvement for agentic workloads — could translate into significantly lower inference costs and higher scalability for deployed applications.
Enterprise teams will likely appreciate the full recipe release, which provides transparency and control that many closed-source solutions cannot match. Organizations with specific compliance, data sovereignty, or customization requirements can now fine-tune and deploy the model within their own infrastructure using NVIDIA’s validated training approach.
The availability across multiple platforms also means teams can experiment with the model using their preferred inference providers before committing to large-scale deployment. This flexibility is particularly important as organizations evaluate different models for agentic use cases where performance characteristics can vary significantly based on specific workflows.
What’s Next
NVIDIA has indicated that the full technical details and benchmarks for Nemotron 3 Super are available in its developer blog posts. The company is expected to continue iterating on the Nemotron family, potentially expanding the hybrid architecture approach to different parameter scales or specialized variants.
For the broader AI ecosystem, the release adds another high-quality open model option at a critical time as developers seek alternatives to rapidly evolving closed-source frontier models. The focus on agentic capabilities suggests NVIDIA will continue investing in architectures and optimizations specifically targeting autonomous AI systems rather than general-purpose chat models.
Developers interested in exploring Nemotron 3 Super can begin testing immediately through the listed platforms, while those wanting to replicate or modify the training process now have access to NVIDIA’s complete methodology.
The move reinforces NVIDIA’s growing role not just as a hardware provider but as an active participant in open-source AI model development, leveraging its computational expertise to address real-world deployment challenges in next-generation AI applications.
Sources
- Introducing Nemotron 3 Super: An Open Hybrid Mamba-Transformer MoE for Agentic Reasoning | NVIDIA Technical Blog
- New NVIDIA Nemotron 3 Super Delivers 5x Higher Throughput for Agentic AI | NVIDIA Blog
- Nvidia launches Nemotron 3 Super AI model for agentic AI systems | Seeking Alpha
- Nvidia launches Nemotron 3 Super, a 120B open model for large-scale AI systems - The New Stack
- NVIDIAAIDev Twitter/X Post