Long-running agent orchestration
Handle the smaller set of difficult planning and synthesis calls inside a multi-model agent system.
Independent tool overview
NVIDIA Nemotron 3 Ultra at a glance
NVIDIA Nemotron 3 Ultra is a frontier-scale, open-weight text model with 550 billion total parameters, 55 billion active parameters, and up to a one-million-token context window for demanding agentic reasoning.
Visit the official NVIDIA Nemotron 3 Ultra site ↗
- Developer
- NVIDIA
- Model size
- 550B total / 55B active parameters
- Context window
- Up to 1 million tokens
- Modalities
- Text input and text output
- Reasoning
- Configurable on or off
- Checkpoints
- BF16, NVFP4, Base BF16, and GenRM
- License
- OpenMDW 1.1
- Release date
- June 4, 2026
Overview
What NVIDIA Nemotron 3 Ultra is
Nemotron 3 Ultra is NVIDIA's largest Nemotron 3 reasoning model, aimed at orchestration, long-horizon agents, coding, tool use, high-stakes RAG, and analysis across very large document or code collections. It is an infrastructure component for developers, not a finished consumer chatbot.
Its hybrid architecture combines Mamba-2, attention, a latent mixture of experts, and multi-token prediction. Only 55 billion of its 550 billion parameters are active for a token, but the complete checkpoint is still enormous: even the smaller NVFP4 release requires a serious multi-GPU deployment.
Developers can try the model through NVIDIA's free API endpoint, download BF16 or NVFP4 weights, or deploy through NVIDIA and open inference tooling. The weights support commercial and non-commercial use under OpenMDW 1.1, but operating cost, safety controls, evaluation, and production reliability remain the deployer's responsibility.
Use cases
Who NVIDIA Nemotron 3 Ultra is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Long-context analysis
Reason across large document sets, repositories, transcripts, or retrieved evidence with up to a 1M-token window.
Agentic coding
Power terminal, repository, debugging, and software-maintenance agents that need multi-step reasoning and tool use.
High-stakes RAG
Act as a reasoning layer over retrieved evidence when the deployment includes rigorous citation and validation controls.
Custom enterprise models
Fine-tune an open-weight frontier model with LoRA, supervised fine-tuning, or reinforcement-learning recipes.
Capabilities
Core NVIDIA Nemotron 3 Ultra features
Hybrid LatentMoE architecture
Combines Mamba-2, mixture-of-experts, and selected attention layers with more efficient latent expert routing.
Multi-token prediction
Predicts multiple future tokens to improve throughput for long outputs and multi-turn agent workflows.
One-million-token context
Supports prompts and outputs within a combined context of up to 1M tokens when the serving stack and memory budget allow it.
Switchable reasoning
The chat template can enable or disable thinking, letting developers reserve deeper reasoning for harder calls.
BF16 and NVFP4 checkpoints
Choose the higher-precision release or the smaller, faster NVFP4 checkpoint based on accuracy, memory, and hardware needs.
Multilingual support
NVIDIA lists English, French, Spanish, Italian, German, Japanese, Hindi, Korean, Brazilian Portuguese, and Chinese.
Open deployment ecosystem
Official recipes cover vLLM, SGLang, TensorRT-LLM, Dynamo, DGX systems, and OpenAI-compatible serving.
Fine-tuning recipes
NVIDIA publishes LoRA, full supervised fine-tuning, and reinforcement-learning paths through NeMo libraries.
Agent-framework examples
Cookbooks connect Ultra to tools such as OpenCode, OpenHands, Hermes Agent, Kilo Code, and Pi.
Process
How the NVIDIA Nemotron 3 Ultra workflow works
Step 1
Prove the task on the hosted endpoint
Build a representative evaluation set and test NVIDIA's API before committing to a multi-GPU deployment.
Step 2
Select the checkpoint
Compare BF16 with NVFP4 on your own tasks; quantization can change performance by benchmark and workload.
Step 3
Choose a serving stack
Use an officially documented vLLM, SGLang, TensorRT-LLM, Dynamo, or NIM route that matches the hardware.
Step 4
Set context and reasoning budgets
Do not default every request to maximum context and open-ended thinking; enforce limits based on task value.
Step 5
Add the agent harness
Connect tools, memory, retrieval, permissions, timeouts, and a sandbox around the model rather than exposing raw actions.
Step 6
Evaluate end to end
Measure task completion, citation fidelity, tool errors, latency, throughput, GPU utilization, and cost on production-like traces.
Step 7
Gate consequential actions
Require deterministic validation or human approval for code execution, external messages, financial operations, and other high-impact steps.
Cost
NVIDIA Nemotron 3 Ultra pricing and free plan
NVIDIA publishes the model weights under OpenMDW 1.1 and currently labels the build.nvidia.com model endpoint free for evaluation. That does not make production inference free: self-hosting requires large GPU systems, while managed deployment, support, and production NIM licensing depend on the provider and infrastructure agreement.
NVIDIA hosted trial
Free endpoint
Try the model and API through build.nvidia.com under NVIDIA's trial terms and service limits.
- Best for evaluation and prototypes
- Account and API key required
- Not a quoted production SLA or unlimited service
Open weights
No model download fee
Download BF16 or NVFP4 weights under OpenMDW 1.1.
- Commercial and non-commercial use supported
- License compliance required
- Compute, storage, networking, and operations are separate
NVFP4 self-hosting
Infrastructure cost
The smaller official checkpoint lowers memory and throughput cost but remains frontier-scale.
- Minimum listed: 4× B200/GB200/GB300/B300 or 8× H100
- Checkpoint is roughly 352 GB on Hugging Face
- Linux and supported NVIDIA serving software required
BF16 self-hosting
Infrastructure cost
Higher-precision deployment with materially greater memory requirements.
- Minimum listed: 8× GB200/B200/GB300/B300, 8× H200, or 16× H100
- Checkpoint is roughly 1.1 TB
- Capacity must also cover KV cache and runtime overhead
Managed production
Provider-specific
Run the model through a cloud inference provider or enterprise NVIDIA deployment.
- Token or GPU-hour pricing varies
- Confirm context, throughput, retention, and SLA terms
- Benchmark total cost on real traces
Pricing checked . Check current pricing at the source ↗
Assessment
NVIDIA Nemotron 3 Ultra strengths and limitations
Where it stands out
- Frontier-scale reasoning with only 55B parameters active per token.
- A one-million-token context window for unusually large reasoning workloads.
- Configurable thinking lets agent systems route easy and hard calls differently.
- Open weights, disclosed data resources, technical reporting, and reproducible evaluation tooling provide unusual transparency for this scale.
- Official BF16 and NVFP4 variants let teams trade precision for footprint and throughput.
- NVIDIA publishes deployment, fine-tuning, and agent-integration recipes across its stack.
What to consider
- Ultra is far beyond workstation scale: the smallest official checkpoint still lists at least four current data-center GPUs or eight H100s.
- The one-million-token limit is a capability ceiling, not a guarantee of accurate retrieval or affordable latency across every token.
- It is text-only; multimodal projects need another model or the separate Nemotron Nano Omni line.
- NVIDIA's benchmark results are informative but do not replace task-specific evaluation, especially because relative BF16 and NVFP4 performance varies by benchmark.
- Open weights do not include a finished agent product, secure sandbox, monitoring, permissions, or human-approval system.
- The model can still hallucinate, follow malicious retrieved instructions, generate vulnerable code, or misuse tools without system-level controls.
- Hardware, interconnect, storage, energy, serving expertise, and low utilization can dominate total cost despite the model having no download fee.
- The OpenMDW license is permissive, but teams still need legal review for their distribution, data, and downstream-use obligations.
Compare
NVIDIA Nemotron 3 Ultra alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Consumer
Nemotron 3 Super
A smaller 120B/12B-active Nemotron with the same 1M context direction and a substantially easier deployment footprint.
Explore Nemotron 3 Super →Consumer
Nemotron 3.5 Lightning
Better for high-volume routine agent execution where speed and efficiency matter more than Ultra-class reasoning.
Explore Nemotron 3.5 Lightning →Coding
GLM-4.7
A smaller open-weight reasoning and coding model with much more accessible self-hosting requirements.
Explore GLM-4.7 →Consumer
Qwen3-Max-Thinking
A competing flagship reasoning model for teams comparing managed frontier intelligence rather than NVIDIA-centered deployment.
Explore Qwen3-Max-Thinking →Questions
NVIDIA Nemotron 3 Ultra FAQs
What is Nemotron 3 Ultra?
It is NVIDIA's frontier-scale open-weight text model for complex reasoning, coding, tool use, long-context analysis, RAG, and long-running agents.
How large is Nemotron 3 Ultra?
The model has 550 billion total parameters and activates about 55 billion parameters per token through its mixture-of-experts architecture.
Is Nemotron 3 Ultra open source?
NVIDIA calls it an open model and releases weights, datasets, recipes, and supporting materials under OpenMDW 1.1. It is safest to describe it as open-weight and review the actual license rather than assuming every artifact uses a standard software open-source license.
Can Nemotron 3 Ultra run locally?
It can be self-hosted, but not on an ordinary PC. NVIDIA lists at least four B200-class GPUs or eight H100s for the NVFP4 checkpoint, with higher requirements for BF16.
How much context does it support?
The official model cards list up to one million tokens. Real usable context depends on serving configuration, KV-cache memory, latency targets, and whether the model accurately uses the relevant evidence.
Is Nemotron 3 Ultra free?
The weights have no download price, and NVIDIA currently provides a free trial endpoint. Production use still incurs GPU, hosting, storage, networking, engineering, support, or managed-provider costs.
What is the difference between BF16 and NVFP4?
BF16 is the larger higher-precision checkpoint. NVFP4 is quantized for a smaller footprint and higher throughput; its accuracy remains close on NVIDIA's published tests but varies above or below BF16 by benchmark.
What is the GenRM variant?
GenRM is a separate Nemotron 3 Ultra–based generative reward model for judging model responses. It is not the default conversational and agentic checkpoint.
Should every agent call use Ultra?
Usually no. NVIDIA positions Ultra for the difficult minority of orchestration and reasoning calls. Smaller Nemotron models can handle routine tool calls and high-volume execution more economically.
Bottom line
Our NVIDIA Nemotron 3 Ultra verdict
Nemotron 3 Ultra is compelling for teams that need open-weight frontier reasoning, huge context, and control over deployment. Its model economics only make sense when difficult agent tasks justify a multi-GPU system; most workloads should route routine steps to smaller models.
Visit NVIDIA Nemotron 3 Ultra website ↗Similar Consumer Tools
View all Consumer AI tools →
MiniMax M3
MiniMax M3 - Open-weight model with 1M context and computer use
Claude Fable
Claude Fable - Anthropic's new frontier Mythos-class model with top performance across benchmarks
Claude Opus 4.8
Claude Opus 4.8 - Anthropic's new top model with improvements to reliability, coding, and agentic flows.
freddy
Freddy - Plug your wearables, CGMs, power meters, and gym apps straight into any AI agent that speaks MCP
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.