NVIDIA Nemotron

Open and efficient multimodal models for agentic AI.

Overview

What Is NVIDIA Nemotron?

NVIDIA Nemotron™ is a family of highly efficient, open, multimodal models, datasets, and technologies built for long-running, specialized agentic AI systems. Designed for advanced reasoning, coding, visual understanding, safety, speech, and information retrieval, Nemotron models help agents complete complex tasks faster with leading accuracy at low cost.

With transparent training data and broad platform support, including NVIDIA RTX PRO™ and NVIDIA DGX Spark™, Nemotron models are openly available and integrated across the AI ecosystem, enabling trusted, high‑performance AI agents to be deployed anywhere from edge to cloud.

NVIDIA Introduces Nemotron 3 Nano Omni for Multimodal Sub-Agents

Meet the new open multimodal model that unifies video, audio, image, and text understanding in a single reasoning loop, replacing fragmented model stacks, and enabling agents to complete tasks faster with less compute.

Nemotron 3 Nano Omni: An Open Omni Reasoning Model for Agentic AI

Learn how Nemotron 3 Nano Omni was designed to deliver leading efficiency and accuracy at low compute. Explore the architecture and training approach, and access open datasets, recipes, cookbooks, and more to get started building agentic AI systems.

Why NVIDIA Built Nemotron

Hear from Bryan Catanzaro, VP of applied deep learning research at NVIDIA, as he shares the vision behind Nemotron and why open technologies are essential for building trusted, enterprise-ready AI.

Benefits

What Does Nemotron Bring to Agentic AI?

Open Models

NVIDIA’s open data and optimization techniques ensure powerful, transparent, and adaptable models for developers and enterprises. Models and training data are published openly on Hugging Face.

High Compute Efficiency

The Nemotron family is optimized to complete agentic tasks faster with the highest throughput and hybrid MoE architecture.

High Accuracy

Built from the ground up with exceptional knowledge, post-trained with high-quality training data, and aligned with reinforcement learning, Nemotron models achieve leading accuracy for long-running agentic systems.

Secure and Simple Deployment

The Nemotron model family, available as optimized NVIDIA NIM™ microservices, offers peak inference performance and flexible deployment options, ensuring superior security, privacy, and portability.

Models

Models for Diverse Workloads

Nemotron models excel in a range of agentic AI tasks, including reasoning, multimodal vision, retrieval-augmented generation (RAG), speech, and safety. Research models are also available for experimentation.

Reasoning

Nemotron models support a range of reasoning workloads. Nano provides superior accuracy and efficiency for specialized sub-agents, Super offers the highest accuracy, throughput reasoning, and tool calling to run complex tasks on multi-agent systems, and Ultra delivers the best reasoning for mission-critical applications that demand maximum capability over multi-step workflows.

Visual Understanding

Multimodal Nemotron models deliver the highest efficiency and leading accuracy across video, audio, image, and text for enterprise agentic use cases. Optimized for specialized sub‑agents, they power capabilities such as computer‑use agents, document intelligence, and video and audio understanding.

Speech

NVIDIA Nemotron Speech models provide high-throughput, ultra-low latency automatic speech recognition (ASR), text-to-speech (TTS), and neural machine translation (NMT) for agentic AI applications.

Retrieval-Augmented Generation

Nemotron RAG models deliver fast, accurate document understanding by extracting multimodal structured information, generating high-quality embeddings, and rank-ordering the most relevant documents. They provide scalable, high-speed retrieval that enhances data quality for LLM training, boosts agent and retriever performance, and streamlines document workflows.

Safety

NVIDIA Nemotron Safety models provide real-time protection against harmful content, off-topic drift, and jailbreak attempts. They add a multilingual, multimodal, content safety layer with reasoning capabilities, enhancing moderation and ensuring cultural alignment.

Technology

Building Blocks for Agentic AI

Start building and optimizing AI agents with NVIDIA NeMo™ for custom agentic AI, NVIDIA NIM for fast, enterprise-ready deployment, and NVIDIA Blueprints for accelerating development with customizable reference workflows.

NVIDIA NeMo

  • Build, customize, and deploy generative AI and agentic AI.
  • Deliver enterprise-ready large language models (LLMs) with precise data curation, cutting-edge customization, scalable data ingestion, RAG, and accelerated performance.
  • Easily build data flywheels and continuously optimize AI agents with the latest information.

NVIDIA NIM

  • Speed up deployment of performance-optimized generative AI models.
  • Run your business applications with stable and secure APIs, backed by enterprise-grade support.

NVIDIA Blueprints

  • Quickly get started with reference applications for generative AI use cases, such as enterprise deep research and multimodal RAG.
  • Accelerate development with blueprints, which include partner microservices, one or more AI agents, reference code, customization documentation, and a Helm chart for deployment.

Starting Options

Ways to Get Started With Nemotron

Start Prototyping for Free

Get started with easy-to-use API endpoints.

  • Access fully accelerated AI infrastructure.
  • Ensure your data isn't used for model training.
  • No credits, just a simple path to build, test, and deploy.

Run Nemotron on Inference Service Providers

Deploy Nemotron models instantly on trusted third-party inference platforms—no infrastructure setup required.

  • Deploy without managing infrastructure.
  • Scale seamlessly from prototype to production.
  • Optimize costs with usage-based pricing.

Get in Touch

Talk to an NVIDIA AI specialist about moving generative AI pilots to production with the security, API stability, and support that comes with NVIDIA AI Enterprise.

Adopters

Enterprises Using Nemotron

Resources

Explore the Latest in Nemotron

  1. Blogs

  2. Sessions

  3. Videos

FAQs

Are NVIDIA Nemotron models open or open source?

NVIDIA Nemotron models aren't just open, but truly open source. NVIDIA publishes the training datasets, techniques, and model weights so the open-source community can benefit from our learnings and use these resources to create their own models.

What does the NVIDIA Open Models License provide to developers?

The NVIDIA Open Model License is a permissive license that allows users to use, modify, distribute, and commercially deploy the models and derivatives without crediting NVIDIA, to encourage innovation and further development of generative AI.

Can I run NVIDIA Nemotron models for free in production?

Yes, you can download and run NVIDIA Nemotron models from Hugging Face for free in production.

Can I run NVIDIA Nemotron models as NVIDIA NIM?

NVIDIA also offers Nemotron models as NVIDIA NIM microservices for secure, scalable deployment, which requires an NVIDIA AI Enterprise license. You can try the Nemotron models and download the NIM microservices from build.nvidia.com.

Is NVIDIA committed to publishing more Nemotron models in the future?

Yes, NVIDIA is committed to publishing more Nemotron models, datasets, and techniques to enable open-source ecosystems.

How do the NVIDIA Nemotron models compare to other open models?

NVIDIA Nemotron models are built on top of frontier open models, making it possible to build better models faster. Additionally, NVIDIA publishes the model weights, training datasets, and training techniques so the developer community can use these different parts of Nemotron to train their own models.

Are the Llama Nemotron models also part of the NVIDIA Nemotron family?

Yes. NVIDIA built the Llama Nemotron models on top of the Llama model family using NVIDIA’s open datasets and advanced techniques, such as Neural Architecture Search (NAS). The Llama Nemotron models inherit the parent Llama model license.[//]: # (The above is a summarization and formatting of the cleaned input.)