Nvidia is shifting its focus towards open infrastructure for artificial intelligence with the introduction of its new Nemotron 3 family of models. These models are designed to help developers create smarter AI agents that can work together, plan, and perform tasks over long periods and across large contexts. Instead of offering ready-made AI services, Nvidia aims to provide tools that allow businesses and engineers to build and customize their own AI solutions more easily.

Introducing Nemotron 3: Flexible and Scalable AI Models

Nemotron 3 comes in three versions: Nano, Super, and Ultra, each tailored for different needs. The Nano is small and efficient, perfect for quick tasks like summarizing content, debugging, or answering questions. It has 30 billion parameters, which activate in groups of 3 billion, enabling fast processing and a context window of one million tokens—meaning it can remember and connect a lot of information during multi-step tasks.

The Super model is much larger, with about 100 billion parameters. It is meant for complex tasks that require multiple AI agents working together, such as deep research or strategic planning. It can activate up to 10 billion parameters per token, offering high accuracy and low latency, making it suitable for demanding applications.

At the top is the Ultra version, with 500 billion parameters and up to 50 billion active per token. This model is designed for highly complex AI applications that need deep reasoning and problem-solving capabilities. All three models use Nvidia’s hybrid latent mixture-of-experts (MoE) architecture, which helps optimize performance and efficiency.

Open Access and Strategic Positioning

Nvidia is making Nemotron 3 Nano available on platforms like Hugging Face and through various inference service providers. It will soon be accessible via Amazon Web Services (AWS) through Amazon Bedrock, and support will expand to Google Cloud, CoreWeave, Microsoft Foundry, and other cloud providers. Additionally, Nvidia offers it as a pre-built NIM microservice, making integration simpler for developers.

The larger models, Super and Ultra, are expected to be released in the first half of 2026. Nvidia emphasizes that these models are not meant to be plug-and-play solutions but rather foundational tools that developers can modify and build upon. This approach positions Nvidia differently from companies like OpenAI or Anthropic, which mainly provide hosted AI services. Instead, Nvidia wants to be the backbone infrastructure that enterprises can use to develop and own their AI agents.

Experts see this as a strategic move to empower businesses with more control over their AI assets. This open approach offers flexibility, enterprise support, and hardware optimization, making it appealing for companies looking to tailor AI to their specific needs rather than relying solely on third-party services.

Hybrid Architecture and Performance Benefits

The core of Nemotron 3’s innovation lies in its hybrid latent mixture-of-experts architecture. This design allows the models to activate only relevant parts of their vast neural networks for each task, which improves speed and reduces resource use. It also enables the models to handle multi-step reasoning and long-term memory, essential features for advanced AI agents that need to operate across extended periods and large datasets.

By offering these models as open, customizable tools, Nvidia is positioning itself as a key player in the AI infrastructure landscape. This move supports the broader trend of enterprises building proprietary AI solutions rather than relying solely on third-party APIs. With flexible models like Nemotron 3, companies can develop more tailored, efficient, and secure AI systems that better fit their workflows and goals.

Overall, Nvidia’s Nemotron 3 aims to provide a robust foundation for the next generation of AI agents, emphasizing openness, scalability, and enterprise readiness. It represents a significant step toward a more collaborative and customizable AI ecosystem, where businesses can innovate freely while leveraging Nvidia’s advanced hardware and software support.

Inspired by

• https://www.computerworld.com/article/4106805/nvidia-bets-on-open-infrastructure-for-the-agentic-ai-era-with-nemotron-3-2.html

Sources

• nvidia.com
• nvidia.com
• huggingface.co
• nvidia.com
• infotech.com