curatedai.net
Directory Categories Radar Models Leaderboard Guides Prompts Skills Briefing Newsletter
Light Dark
← Back
← All NVIDIA models
LLMS • CURATED • UPDATED AUG 4, 2026

NVIDIA Nemotron 3 Ultra

NVIDIA's 550B open-weights reasoning model, built for inference speed

Nemotron 3 Ultra is the largest member of NVIDIA's Nemotron 3 family, released 4 June 2026 at Computex. It is a 550B-parameter hybrid latent mixture of experts with roughly 55B parameters active per token, combining Mamba and Transformer blocks, and trained in NVIDIA's 4-bit NVFP4 format on Blackwell hardware. It ships under the NVIDIA Open Model License, which permits commercial use, and NVIDIA published training data, reinforcement learning environments and post-training recipes alongside the weights rather than the weights alone. Weights are on Hugging Face as nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B in both BF16 and NVFP4, and it is served through OpenRouter, Together AI, Baseten, DeepInfra, Fireworks and NVIDIA NIM.

QUICK FACTS

Pricing Free
Platforms api, local
API Yes
Open Source Yes
Modalities LLMs
Best For Best for Open-Weight Throughput
Date Added 2026-08-04

WHY WE PICKED IT

The architecture is optimised for throughput rather than peak benchmark score, and it shows: over 400 output tokens per second at 550B parameters. It is also the most openly documented release at this scale, because publishing the datasets and post-training recipes lets you actually reproduce and extend the model instead of just running it.

STRENGTHS

LIMITATIONS

GETTING STARTED

The fastest way in is OpenRouter or Together AI, where it is already served. To self-host, take the NVFP4 checkpoint if you are on Blackwell and BF16 otherwise. Read the NVIDIA Open Model License before building on it commercially, because it is NVIDIA's own licence rather than Apache or MIT, and the terms are worth checking against your use case.

QUICK TIPS

1 Choose NVFP4 over BF16 on Blackwell, that is what the training format was for
2 Benchmark tokens per second, not just quality, because throughput is this model's actual advantage
3 Compare against Inkling and GLM-5.2 before committing, they score higher on the independent index
4 Use the published RL environments if you intend to post-train it yourself
5 Read the licence terms; open weights and open source are not the same thing here

SIMILAR TOOLS

Claude Fable 5 NotebookLM Claude Opus 5 GPT-5.6 Sol Kimi K3

USE CASE EXAMPLES

Latency-Sensitive Agent Serving

Serve an agent loop where response speed matters more than peak reasoning.

STEPS:
  1. Deploy the NVFP4 checkpoint on Blackwell hardware
  2. Measure tokens per second under your real concurrency
  3. Set a token budget per agent step
  4. Route only the hardest steps to a stronger model
  5. Track cost per completed task against a hosted frontier model
  6. Scale replicas rather than model size when throughput is the constraint

Post-Training On Published Recipes

Extend the model using the same tooling NVIDIA used to build it.

STEPS:
  1. Download the weights plus the published post-training recipes
  2. Set up NeMo RL and NeMo Evaluator
  3. Reproduce a baseline evaluation before changing anything
  4. Introduce your domain data or RL environment
  5. Evaluate against the untouched checkpoint
  6. Publish or deploy only if the delta holds on held-out tasks

PRICING

Free Completely free
📚

FEATURED IN GUIDES

How To Run AI Models Locally: Hardware, Quantisation And Engines

Running a frontier open model locally is a memory problem before it is a speed problem, and a bandwi...

Uncensored AI Models: What Abliteration Actually Does To Them

Abliteration removes a model's ability to refuse by deleting one direction from its activations. It ...

Which LLMs Actually Deliver in 2026?

Comprehensive comparison of the best large language models in 2026 including ChatGPT, Claude, Gemini...

Large Language Models Explained: How They Actually Work

Complete guide to Large Language Models (LLMs). Learn how LLMs work, their architecture, training pr...

Choosing the Right LLM: Decision Framework That Actually Works

Complete guide to choosing the right large language model for your needs. Compare ChatGPT, Claude, G...

EXPLORE ALTERNATIVES

View NVIDIA Nemotron 3 Ultra Alternatives (2026) →

Compare NVIDIA Nemotron 3 Ultra with 5+ similar llms AI tools.

❓

FREQUENTLY ASKED QUESTIONS

Q

Is NVIDIA Nemotron 3 Ultra free?

A

Yes, NVIDIA Nemotron 3 Ultra is completely free to use.

Q

Does NVIDIA Nemotron 3 Ultra have an API?

A

Yes, NVIDIA Nemotron 3 Ultra offers an API for programmatic integration.

Q

What is NVIDIA Nemotron 3 Ultra best for?

A

NVIDIA Nemotron 3 Ultra is best for Best for Open-Weight Throughput. Nemotron 3 Ultra is the largest member of NVIDIA's Nemotron 3 family, released 4 June 2026 at Computex. The architecture is optimised for throughput rather than peak benchmark score, and it shows: over 400 output tokens per second at 550B parameters. It is also the most openly documented release at this scale, because publishing the datasets and post-training recipes lets you actually reproduce and extend the model instead of just running it.

Q

What platforms does NVIDIA Nemotron 3 Ultra support?

A

NVIDIA Nemotron 3 Ultra supports api, local.

Q

Is NVIDIA Nemotron 3 Ultra open source?

A

Yes, NVIDIA Nemotron 3 Ultra is open source. You can access the source code, contribute, and deploy it on your own infrastructure.

Q

How do I get started with NVIDIA Nemotron 3 Ultra?

A

The fastest way in is OpenRouter or Together AI, where it is already served. To self-host, take the NVFP4 checkpoint if you are on Blackwell and BF16 otherwise. Read the NVIDIA Open Model License before building on it commercially, because it is NVIDIA's own licence rather than Apache or MIT, and th...

Q

How do I use NVIDIA Nemotron 3 Ultra?

A

NVIDIA Nemotron 3 Ultra is a large language model for text generation, analysis, and conversation. Use the API for programmatic access. Enter prompts or questions to get responses. It excels at over 400 output tokens per second despite 550b total parameters.

🏷️

FEATURED ON CURATEDAI

Work on NVIDIA Nemotron 3 Ultra? You're hand-reviewed in our directory. Add this badge to your site — it links back to this profile.

Featured on CuratedAI