Models·NVIDIA

Nemotron 3 Ultra: Review, Pricing, and Limits

nvidia/nemotron-3-ultra-550b-a55b

Nemotron 3 Ultra is NVIDIA’s open-weight text model for reasoning, orchestration, and tool use.

$0.50Input, per million tokens
$2.20Output, per million tokens
262KContext window
16KLongest answer

262K tokens of context — about 196,608 words, on a scale that ends at 2M.

What it can do

  • Reasoning
  • Tool use
  • Open weights
  • Takes text
  • Gives text

Released 4 Jun 2026 · checked 3 Oct 2026.

About it

Nemotron 3 Ultra is an NVIDIA text model released in June 2026. It uses a mixture-of-experts design with 55 billion active parameters from 550 billion total, supports reasoning and tools, and has open weights. It can read very large text inputs, but it only handles text and has a fixed maximum answer length.

Review

In a line

Nemotron 3 Ultra is an open-weight language model from NVIDIA for text reasoning and orchestration. It accepts text and produces text, with support for tools.

What it's good at

Use it for tasks that need step-by-step reasoning, coordinating actions, or calling tools. Its design activates 55 billion of 550 billion total parameters at a time, using a mixture-of-experts approach.

Price in plain terms

Input costs $0.5 per million tokens, and output costs $2.2 per million tokens. A token is a small piece of text counted for billing.

Reading a roughly 500-word page costs about 0.033 cents. A chat turn with about 1,000 tokens in and 500 out costs about 0.16 cents. One thousand such turns cost about $1.60.

Context and limits

The model can take up to 262,144 tokens in one request, so it can work with very long text. Each answer can contain up to 16,384 tokens.

It is text-only: both its inputs and outputs are text.

Ways to use it

Use Nemotron 3 Ultra in a service or app that offers the model under its NVIDIA model ID. It can return text and use supported tools.

Its weights are open, so teams looking through open-weight models can consider it for setups that support this model.

Verdict

Nemotron 3 Ultra fits readers who need text reasoning, orchestration, and tool calls, especially when long documents must fit in one request. Its open weights may suit teams choosing from open-weight models.

Skip it if you need images or another input or output format. Also check that your chosen app or service supports the model ID and its 262,144-token context before relying on it.

Verdict

Choose Nemotron 3 Ultra for text-heavy work where reasoning, tool use, or coordinating steps matters. Its large context can hold substantial material, and the per-token rates make many ordinary chat turns inexpensive. Open weights give it a place in model selection for teams with a compatible setup. It is not a general media model: the listed inputs and outputs are text only. Confirm that your software supports this specific NVIDIA model before committing to it.

  • Open weights are available.
  • Supports reasoning and tool use.
  • 262,144-token context window.
  • About $1.60 for 1,000 typical chat turns.
  • Text input and output only.
  • Maximum response is 16,384 tokens.
  • Requires a compatible service or setup for its model ID.
  • Pick it ifLong text analysis and reasoning; Tool-based orchestration; Teams considering open-weight models
  • Skip it ifYou need image or non-text input and output.; Your app or provider does not support Nemotron 3 Ultra.

Pricing

Input is $0.5 per million tokens and output is $2.2 per million tokens. Typical costs are about 0.033 cents to read a page, 0.16 cents per chat turn, and $1.60 for 1,000 such turns.

Questions

Who makes Nemotron 3 Ultra?

NVIDIA makes Nemotron 3 Ultra. It belongs to NVIDIA’s model family.

Is Nemotron 3 Ultra open source?

Its weights are open. The listed model ID is nvidia/nemotron-3-ultra-550b-a55b.

What can Nemotron 3 Ultra do?

It handles text, supports reasoning, and can use tools for orchestration tasks.

How much text can it read?

Its context window is 262,144 tokens, meaning that is the maximum input size for one request.

How long can its answer be?

The maximum output is 16,384 tokens.

What does it cost?

Input is $0.5 per million tokens and output is $2.2 per million tokens. A typical 1,000-turn example costs about $1.60.

Can it handle images?

No image capability is listed. Its inputs and outputs are text.

On these boards

More from NVIDIA