NVIDIA Nemotron 3 Ultra
- Score
- 11.8 · rank 36
- Price
- Free
- API
- Yes
- Weights
- Yes
WHAT IT DOES
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.
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.
QUICK FACTS
STRENGTHS
- ✓Over 400 output tokens per second despite 550B total parameters
- ✓NVIDIA Open Model License permits commercial use
- ✓Training data, RL environments and post-training recipes published with the weights
- ✓Both BF16 and NVFP4 checkpoints, so it runs efficiently on Blackwell
- ✓Served by six or more inference providers, not just NVIDIA's own
LIMITATIONS
- ⚠Distributed as code and model weights; there is no hosted product to sign into, so running it means self-hosting or a third-party host.
- ⚠At 37.8 on the current Artificial Analysis index it sits well behind GLM-5.2 and Inkling, and far behind the frontier
- ⚠The NVIDIA Open Model License is not OSI open source, so the terms need reading before commercial use
- ⚠550B parameters means the hardware bill is real even at 55B active
- ⚠Launch coverage quoted 47.7 on the index; that was the pre-v4.1 scale and no longer compares to current numbers
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.
BENCHMARKS
QUICK TIPS
OFFICIAL LINKS
SIMILAR TOOLS
USE CASE EXAMPLES
Latency-Sensitive Agent Serving
Serve an agent loop where response speed matters more than peak reasoning.
- Deploy the NVFP4 checkpoint on Blackwell hardware
- Measure tokens per second under your real concurrency
- Set a token budget per agent step
- Route only the hardest steps to a stronger model
- Track cost per completed task against a hosted frontier model
- 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.
- Download the weights plus the published post-training recipes
- Set up NeMo RL and NeMo Evaluator
- Reproduce a baseline evaluation before changing anything
- Introduce your domain data or RL environment
- Evaluate against the untouched checkpoint
- Publish or deploy only if the delta holds on held-out tasks
PRICING
FEATURED IN GUIDES
When to Download Mistral Large 4
Mistral published the weights for Large 4, nicknamed Le Chonk, on October 29. What the licence allow...
Who Can Use Gemini 4 Carbon
Google announced Gemini 4 Carbon on October 28 for a small set of cyber defenders. Who has access, w...
When to Use GPT-6.1 Terra
OpenAI priced GPT-6.1 Terra at $1 per million input tokens for computer use and everyday writing. Ho...
Frontier Training Without the Live Web
OpenAI says frontier training will use cached and simulated web content, after a model reached the l...
Open Models: What They Cannot Do
The benchmark gap has largely closed: open models now score within a point of the closed frontier on...
EXPLORE ALTERNATIVES
Compare NVIDIA Nemotron 3 Ultra with 5+ similar llms AI tools.
FREQUENTLY ASKED QUESTIONS
Is NVIDIA Nemotron 3 Ultra free?
Yes, NVIDIA Nemotron 3 Ultra is completely free to use.
Does NVIDIA Nemotron 3 Ultra have an API?
Yes, NVIDIA Nemotron 3 Ultra offers an API for programmatic integration.
What is NVIDIA Nemotron 3 Ultra best for?
NVIDIA Nemotron 3 Ultra is 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.
What platforms does NVIDIA Nemotron 3 Ultra support?
NVIDIA Nemotron 3 Ultra supports api, local.
Is NVIDIA Nemotron 3 Ultra open source?
Yes, NVIDIA Nemotron 3 Ultra is open source. You can access the source code, contribute, and deploy it on your own infrastructure.
How do I get started with NVIDIA Nemotron 3 Ultra?
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...
How do I use NVIDIA Nemotron 3 Ultra?
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.
