NVIDIA Nemotron 3 Ultra
AI model · Language model (LLM) · Text → Text by Nvidia
Frontier intelligence for complex agentic AI applications.
NVIDIA Nemotron 3 Ultra is a large language model with 550 billion parameters, offering frontier-level intelligence for complex agentic AI applications. It features a 1M token context window, high inference throughput, and multilingual support, making it ideal for advanced reasoning and automation tasks.
Category: Generative AI
- Context window
- 1M tokens
- Parameters
- 550B total, 55B active
- Licence
- OpenMDW License Agreement, version 1.1
- Released
- 2026-06-04
- Total Parameters
- 550B
- Active Parameters
- 55B
- 550 billion total parameters with 55 billion active parameters
- 1M token context window for extensive analysis
- Up to 6x higher inference throughput than state-of-the-art LLMs
- Supports multilingual reasoning in 11 languages
- Open-sourced checkpoints, data, and recipes
Overview
NVIDIA Nemotron 3 Ultra is a powerful large language model designed for complex agentic AI applications, offering frontier-level intelligence and high throughput. It is suitable for businesses and developers looking to build sophisticated AI agents that can reason, plan, and act autonomously. Nemotron 3 Ultra features approximately 550 billion total parameters and 55 billion active parameters, with a remarkable context window of up to 1 million tokens. This allows for deep analysis of extensive information and long-running tasks. The model achieves up to 6x higher inference throughput compared to state-of-the-art open LLMs, enabling faster task completion and reduced operational costs. It supports multilingual reasoning across 11 languages and is available under the OpenMDW License Agreement, version 1.1, with open-sourced checkpoints, data, and recipes.
Capabilities
- Reasoning engine: Nemotron 3 Ultra is a large reasoning engine designed for complex agentic AI applications.
- Frontier intelligence: It offers frontier-level intelligence and high throughput for AI agents.
- 1M token context window: It can process up to 1 million tokens in its context window.
- High inference throughput: It achieves up to 6x higher inference throughput compared to state-of-the-art publicly available LLMs.
- Agent optimization: It is optimized for leading agent platforms and harnesses, enabling faster task completion and lower operational costs for AI agents.
- Multilingual support: It supports multilingual reasoning across English, French, Spanish, Italian, German, Japanese, Korean, Hindi, Brazilian Portuguese, and Chinese.
- Use cases: It can be used for coding assistants, search, complex workflow automation, and high-stakes RAG.
- Configurable reasoning mode: It can be configured to enable or disable its reasoning mode via the chat template.
Technical details
- Architecture: The model uses a Mixture-of-Experts Hybrid Mamba-Attention architecture with LatentMoE and Multi-Token Prediction (MTP).
- Training methods: It is pretrained in NVFP4 and post-trained with Supervised Fine Tuning (SFT), Reinforcement Learning (RL), and Multi-teacher On-Policy Distillation (MOPD).
- Reasoning budget control: It supports reasoning budget control.
Use cases
- Building advanced AI agents
- Enabling complex reasoning and decision-making
- Automating coding, search, and workflow tasks
- Processing and analyzing long-context documents
- Developing multilingual AI applications
- High-stakes Retrieval-Augmented Generation (RAG)
Specifications
- Context window
- 1M tokens
- Parameters
- 550B total, 55B active
- Licence
- OpenMDW License Agreement, version 1.1
- Released
- 2026-06-04
- Languages
- English, French, Spanish, Italian, German, Japanese, Hindi, Korean, Brazilian Portuguese, Chinese
- Total Parameters
- 550B
- Active Parameters
- 55B
- Architecture
- Mixture-of-Experts Hybrid Mamba-Attention with LatentMoE and MTP
- Context Length
- Up to 1M tokens
- Inference Throughput
- Up to 6x higher than state-of-the-art LLMs
- Training Data (Pretraining)
- 20 trillion text tokens
- Post-training methods
- SFT, RL, MOPD
- Quantization
- NVFP4 (available)
- License
- OpenMDW License Agreement, version 1.1
- Release Date
- 2026-06-04
How it works
- Deployment options: Access the model through NVIDIA NIM microservices for enterprise-ready deployment, or use open frameworks like vLLM, SGLang, Ollama, and llama.cpp on NVIDIA GPUs.
- Fine-tuning with NeMo AutoModel: Fine-tune the model using NVIDIA NeMo AutoModel with provided recipes for PEFT (LoRA) on datasets like HellaSwag.
- Leveraging for agentic AI: Utilize the model for complex agentic workflows, coding assistance, search, and multilingual reasoning tasks.
FAQ
What is Nemotron 3 Ultra?
NVIDIA Nemotron 3 Ultra is a large language model with approximately 550 billion total parameters and 55 billion active parameters. It utilizes a hybrid Mamba-Attention architecture with Mixture-of-Experts (MoE) and LatentMoE for enhanced performance.
What are the primary use cases for Nemotron 3 Ultra?
Nemotron 3 Ultra is designed for complex agentic AI applications, offering frontier-level intelligence, high inference throughput, and a 1M token context window. It is ideal for tasks such as coding assistance, search, workflow automation, and multilingual reasoning.
What languages does Nemotron 3 Ultra support?
Nemotron 3 Ultra supports a wide range of languages including English, French, Spanish, Italian, German, Japanese, Hindi, Korean, Brazilian Portuguese, and Chinese.
What is the licensing for Nemotron 3 Ultra?
The model is released under the OpenMDW License Agreement, version 1.1, and its checkpoints, training data, and recipes are open-sourced.