Skip to content
ShilpBot
AI Models

NVIDIA Nemotron 3 Ultra (Free): Specs, Limits and How to Use It

NVIDIA Nemotron 3 Ultra free: hybrid Transformer-Mamba mixture-of-experts model, text input, 1M context. Specs, limits, test prompts and how to try it.

By ShilpBot Team · Published · 3 min read

Nemotron 3 Ultra is NVIDIA's open frontier-reasoning and orchestration model, with 55 billion active parameters out of 550 billion total. It uses a hybrid Transformer-Mamba mixture-of-experts architecture.

It is the largest of the free NVIDIA models in this series and the one with the most headroom for hard reasoning. It exposes a reasoning effort setting, so you can trade speed for depth on a per-request basis.

Specifications checked against the public OpenRouter catalogue on 2026-11-16. See the live model page for current limits.

NVIDIA Nemotron 3 Ultra quick facts

DetailValue
DeveloperNVIDIA
Model IDnvidia/nemotron-3-ultra-550b-a55b:free
Added to the catalogue2026-06-04
TypeHybrid Transformer-Mamba mixture-of-experts model
Parameters55B active / 550B total
Context windowabout 1M tokens (1,000,000)
Maximum output65,536 tokens
Input typestext
PriceFree (paid listing: $0.60 in / $2.40 out per 1M tokens)

Feature support from the catalogue: Tool calling: listed; JSON / structured output: not listed; Adjustable reasoning effort: listed.

What NVIDIA Nemotron 3 Ultra is good for

  • Hard multi-step reasoning, planning and orchestration of other tools or agents.
  • Long-context analysis up to 1M tokens.
  • Cases where you want to compare an open model against closed frontier models.

Limits of NVIDIA Nemotron 3 Ultra

  • Larger models are slower per token. Use a lighter model for simple tasks.
  • Text only, so use another model for images.
  • Free tiers may use your prompts for training, as the notice in many model pickers says.

How NVIDIA Nemotron 3 Ultra compares

  • NVIDIA Nemotron 3.5 Lightning: Both have a 1,000,000-token context window; both cap output at 65,536 tokens; it activates 55B parameters per token against 3B.
  • Baidu ERNIE 5.5 Flash Preview: It has a smaller context window (1,000,000 against 1,024,000 tokens); its output cap is 65,536 tokens against 921,600.
  • Thinking Machines Inkling 2 Small: It has a smaller context window (1,000,000 against 1,048,576 tokens); its output cap is 65,536 tokens against 131,072; it does not accept image input; it activates 55B parameters per token against 17B.

Calling NVIDIA Nemotron 3 Ultra from code

Use the model ID nvidia/nemotron-3-ultra-550b-a55b:free with any OpenAI-compatible client pointed at OpenRouter:

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "nvidia/nemotron-3-ultra-550b-a55b:free", "messages": [{"role": "user", "content": "Explain what you can do in two sentences."}]}'

Prompts that show what NVIDIA Nemotron 3 Ultra can do

  1. Give it a project description and ask it to break the work into tasks for three specialised agents.
  2. Ask for a step-by-step proof or derivation and check each step.
  3. Compare low and high reasoning effort on the same problem, noting time and correctness.

What NVIDIA Nemotron 3 Ultra costs

The free listing costs nothing, but the same model also has a paid listing at about $0.60 per million input tokens and $2.40 per million output tokens. As a rough guide, sending 1 million input tokens and receiving 100,000 output tokens would cost about $0.84 on the paid listing. Use the free tier to test, and move to the paid route if you need steadier availability. Free pricing is set by the provider and can change at any time, so check the model page before you plan around it.

Keep reading

All free models are compared in our free AI chat models roundup. New to terms like mixture-of-experts? Start with how neural networks learn.

Frequently asked questions

What does a hybrid Transformer-Mamba architecture mean?

It combines Transformer attention layers with Mamba state-space layers, which handle long sequences efficiently, inside a mixture-of-experts design.

Is NVIDIA Nemotron 3 Ultra free, and what are its limits?

As of 2026-11-16 the catalogue lists it at zero price. About 1M tokens (1,000,000) context window, up to 65,536 output tokens, model ID nvidia/nemotron-3-ultra-550b-a55b:free. Free use can be rate limited and providers may change it.

#NVIDIA Nemotron 3 Ultra#NVIDIA free model#free AI model#AI chat model#reasoning model#Kilo Gateway#OpenRouter