MODELS
Nemotron 3 Ultra
by NVIDIA
Overview
NVIDIA Nemotron 3 Ultra is an open 550B-parameter MoE language model with 55B active parameters, 1M context, configurable reasoning, and tool-use support.
Details
NVIDIA Research presents Nemotron 3 Ultra as its most capable Nemotron 3 model, with 550B total parameters and 55B active parameters. NVIDIA’s launch blog describes it as an open MoE model for reasoning, orchestration, long-running agents, coding, tool use, and long-context work. The NVIDIA NIM model card lists text input/output, 1M context, configurable reasoning, tool use, supported languages, OpenMDW-1.1 license, and a June 4, 2026 release date. NVIDIA API documentation provides the nvidia/nemotron-3-ultra-550b-a55b chat-completions endpoint.
When to Use
Evaluate NVIDIA’s highest-capability Nemotron 3 model profile for large-scale reasoning and long-context workloads. Build agentic workflows that need reasoning orchestration coding tool use or long-running agent behavior as described by NVIDIA. Use an NVIDIA NIM/API-hosted model with official model card API reference deployment guide and downloadable checkpoint coverage.
Getting Started
- Read the NVIDIA Research Nemotron 3 Ultra page and technical report for the model overview.
- Review the NVIDIA Build/NIM model card for capabilities
- context length
- license
- and release metadata.
- Use the NVIDIA API reference for the nvidia/nemotron-3-ultra-550b-a55b chat-completions endpoint.
- Follow the NVIDIA NIM deployment guide for OpenAI-compatible and Anthropic-compatible APIs
- tool calling
- reasoning controls
- and agentic workflows.
- Check NVIDIA Developer Nemotron resources for the paper
- Hugging Face collection
- datasets
- and GitHub deployment cookbooks.
Key Features
- •550B total parameters with 55B active parameters.
- •Open MoE model
- •according to NVIDIA’s launch blog.
- •1M context listed on the NVIDIA NIM model card and API/catalog pages.
- •Configurable reasoning and tool-use support listed on the NVIDIA NIM model card.
- •NIM exposes OpenAI-compatible and Anthropic-compatible APIs for chat completions
- •responses
- •messages
- •tool calling
- •reasoning controls
- •and agentic workflows
- •according to NVIDIA Docs.
- •Official NVIDIA resources include a research page
- •launch blog
- •NIM model card
- •API reference
- •technical report
- •Hugging Face model card
- •datasets
- •and deployment cookbooks.
Capabilities
- •text-input-output
- •chat-completions
- •long-context
- •reasoning
- •configurable-reasoning
- •tool-use
- •coding
- •agent-orchestration
- •openai-compatible-api
- •anthropic-compatible-api
- •model-checkpoints
- •deployment-resources
Last updated Jul 31, 2026