NVIDIA has officially unveiled its latest family of open models, Nemotron 3, promising enhanced AI performance across a spectrum of applications. With three distinct model sizes designed for efficiency and scale, Nemotron 3 is setting new benchmarks in the AI industry.
NVIDIA Introduces Nemotron 3 AI Models
In a recent announcement, NVIDIA introduced the Nemotron 3 family, an innovative suite of open models, data, and libraries. These models are aimed at transforming agentic AI development, offering transparency and efficiency across various sectors. The standout feature of Nemotron 3 is its unique hybrid latent mixture-of-experts (MoE) architecture, which facilitates the creation and deployment of reliable multi-agent systems on a large scale.
Designed to support NVIDIA’s sovereign AI initiatives, Nemotron 3 models are gaining traction among organizations worldwide, including those in Europe and South Korea. This adoption underscores the models’ ability to align AI systems with local data, regulations, and values, enhancing their global appeal.
Among the early adopters of Nemotron 3 are industry giants like Accenture, Deloitte, Oracle Cloud Infrastructure, Siemens, and Zoom. These companies are leveraging the models to enhance AI workflows across diverse fields such as manufacturing, cybersecurity, and media.
Nemotron 3: A Closer Look at the Models
The Nemotron 3 family consists of three models tailored for different levels of AI tasks:
- Nemotron 3 Nano: This model features 30 billion parameters with 3 billion active, tailored for highly efficient and specific tasks.
- Nemotron 3 Super: Approximately 100 billion parameters with 10 billion active, ideal for applications requiring high-accuracy multi-agent collaboration.
- Nemotron 3 Ultra: With around 500 billion parameters and 50 billion active, this model is perfect for complex AI applications demanding deep research and strategic planning.
Nemotron 3 reinvents multi-agent AI with a focus on efficiency and accuracy, achieving up to 4x higher token throughput and 60% reduction in token generation costs.
Nemotron 3 Nano is acclaimed for its cost-efficient design, supporting tasks like software debugging and content summarization. The model’s architecture allows for significant gains in efficiency and scalability, making it a favored choice for AI development.
Artificial Analysis has recognized Nemotron 3 Nano for its openness, efficiency, and leading accuracy among its peers. The Super and Ultra models utilize NVIDIA’s 4-bit NVFP4 training format, offering reduced memory requirements and faster training, making them a viable option for extensive infrastructure without sacrificing accuracy.
Availability and Deployment
Developers can access Nemotron 3 Nano today via Hugging Face and numerous inference service providers. The model is also available on various enterprise AI platforms, with forthcoming availability on major cloud platforms like AWS and Google Cloud.
With Nemotron 3 anticipated to advance further in 2026, including the release of the Super and Ultra models, NVIDIA continues to provide scalable, secure AI solutions tailored for diverse industry needs.