NVIDIA Nemotron 3 Ultra: 550B Open Model for Agentic AI

On June 4, 2026, NVIDIA announced Nemotron 3 Ultra, the biggest model in the Nemotron 3 family. It comes with 550 billion total parameters, although only 55 billion are active at a time through its Mixture-of-Experts architecture.

The target here is not ordinary chat.

Nemotron 3 Ultra is designed around agentic workloads: systems that can run for a long time, use tools, write code, perform research, and coordinate multiple agents to complete a larger task.

The other important part is the way NVIDIA is releasing it. The model is fully open-source, with the weights, training recipes, and datasets available. This is not simply another model that you access through a hosted API. The idea is to provide the underlying infrastructure as well.

Model Specifications

Why This Matters for Agentic AI

An agentic system has a different set of requirements from a chatbot.

A normal chat session might only need enough context to answer the current conversation. An agent can accumulate tool results, files, previous decisions, intermediate reasoning, and other state over hundreds or thousands of steps.

That creates four fairly obvious requirements:

• Massive context — agents accumulate history, tool outputs, and intermediate reasoning across thousands of steps
• High throughput — inference cost compounds when agents run for hours
• Reliable tool calling — structured output, function calling, code execution
• Reasoning control — budget-aware inference for cost/quality tradeoffs

Nemotron 3 Ultra is built around those requirements.

Its 1M-token context window outperforms state-of-the-art open LLMs on RULER at the 1M-token length. NVIDIA's hybrid Mamba-Attention architecture with LatentMoE is also aimed at keeping inference efficient. On 8k input / 64k output settings, NVIDIA reports 5.9x higher inference throughput than GLM-5.1-754B-A40B and 4.8x higher throughput than Kimi-K2.6-1T-A32B.

The two Multi-Token Prediction (MTP) layers are useful here as well. They enable native speculative decoding, while the reasoning budget controls give developers a way to put a limit on how much compute the model can spend on each step.

For long-running agents, that kind of control matters. A system that can reason indefinitely is not necessarily useful if every step becomes too expensive.

Free API Access (OpenRouter)

If you want to try Nemotron 3 Ultra without setting up a datacenter first, OpenRouter currently provides a free route:

Model ID: nvidia/nemotron-3-ultra-550b-a55b:free
Context: 1,000,000 tokens
Max Output: 65,536 tokens
Rate Limits (free tier): 20 RPM, 200 RPD

Example API call:

from openai import OpenAI

client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key="YOUR_OPENROUTER_KEY"
)

response = client.chat.completions.create(
    model="nvidia/nemotron-3-ultra-550b-a55b:free",
    messages=[{"role": "user", "content": "Design a multi-agent research system"}],
    max_tokens=8192
)

The free tier is limited to 20 requests per minute and 200 requests per day. If that is not enough, adding $5–10 to the account unlocks the "Paid tier", which raises the allowance to 1,000 requests per day. The limit remains at 20 requests per minute, while paid models are not subject to OpenRouter's free-tier limits.

That makes the free endpoint useful for testing and prototyping, even if it is not necessarily how you would run a production agent system.

Self-Hosted Deployment (Unlimited)

For production workloads, the other option is to run the model yourself. Self-hosting removes the API limits and gives you control over the inference stack.

NVIDIA also provides several checkpoints through Hugging Face:

• nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4 — quantized, post-trained
• nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 — post-trained
• nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-Base-BF16 — base model
• nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-GenRM — reward model for RLHF

There is one fairly important catch: 550 billion parameters is still a lot of model to run.

This is multi-GPU territory, with H100 or A100-class hardware recommended. NVFP4 makes the model more manageable, but it does not turn a 550B model into something that can casually run on a desktop GPU.

If you only have a single GPU, NVIDIA's smaller models make more sense. Nemotron 3 Super is 120B/12B, while Nemotron 3.5 Lightning is 30B/3B with MTP.

Training Pipeline (Fully Reproducible)

One of the more interesting parts of the release is that NVIDIA did not stop at publishing the model weights. The complete training recipe is available in NVIDIA-NeMo/Nemotron.

The pipeline looks like this:

Pretrain (20T tokens, NVFP4, 2-phase curriculum)
    ↓
SFT — multi-domain, 12+ data sources
    ↓
RLVR — 21 verifiable reward environments
    ↓
MOPD — Multi-teacher On-Policy Distillation
    ↓
MTP Boosting — Multi-Token Prediction heads

NVIDIA also released several of the datasets used in the process:

• Nemotron-Pretraining-Code-v3 — 173B tokens of fresh GitHub code (through Sept 2025)
• Nemotron-Pretraining-Legal-v1 — synthetic legal reasoning
• Nemotron-Pretraining-Specialized-v1.2 — factual recall, moral scenarios, MCQ/generative QA
• Nemotron-Posttraining-v3 — agentic, reasoning, general capability SFT/RL data

That makes the release more useful to researchers and teams that want to experiment with the training approach itself, rather than simply downloading a checkpoint and running inference.

The Nemotron 3 Family (Right-Size Your Agent)

Ultra is the flagship model, but it is not the only option in the family.

That range is important because not every agent needs a 550B model.

Nano and Lightning are much more practical for edge or high-volume workloads. Super sits somewhere in the middle for teams that want stronger reasoning and coding without moving all the way into datacenter-scale Ultra deployment.

Ultra is really aimed at the top end: large, long-running agent systems where the additional capability and context are worth the infrastructure cost.

What This Enables

With Nemotron 3 Ultra, there are several workloads that become particularly interesting:

• Research agents that ingest 1M-token corpora and synthesize findings
• Coding agents that maintain full repository context across thousands of edits
• Multi-agent orchestration with shared long-term memory
• Autonomous workflows that run for hours with budget-controlled reasoning

The 1M-token context is especially useful for applications where repeatedly summarising or trimming context would otherwise become a problem. A research agent, for example, can work with a much larger body of source material without constantly throwing older information away.

For coding agents, the same idea applies to large repositories and long-running tasks. Instead of repeatedly rebuilding context around a smaller window, the system can retain much more of the project state.

The open training recipes are another part of the story. Teams can replicate the methodology, adapt parts of it, or use their own data where the license and training setup allow it.

The NVIDIA Nemotron Open Model License also permits commercial use and on-prem deployment, which makes the model relevant to organisations that cannot or do not want to send their workloads to an external API.

Bottom Line

Nemotron 3 Ultra is NVIDIA's flagship open model for agentic workloads: 550 billion total parameters, 55 billion active, a 1M-token context window, and a hybrid Mamba-Attention MoE architecture.

More importantly, NVIDIA is releasing the weights alongside the training recipes and datasets. That makes it a piece of open infrastructure rather than just another model available through an API.

For experimenting, the OpenRouter free tier is probably the easiest place to start. For production workloads, the more serious route is self-hosting with something like TensorRT-LLM and NVFP4, assuming you have the GPU capacity to support a model of this size.

And if Ultra is simply too large for the hardware you have, the rest of the Nemotron 3 family gives you smaller options.

Agentic systems are going to need models that can keep context, use tools, reason through long tasks, and run without inference costs getting completely out of hand. Nemotron 3 Ultra is NVIDIA's attempt to put all of those pieces into one open model.