NewChat Models

NVIDIA

Nemotron 3 Ultra

NVIDIA’s flagship Nemotron reasoner for complex text, code and multi-step analysis, using a paid 256K OpenRouter route.

Reasoning256K contextText
Open AppBrowse All Models

At a glance

Know the model before you prompt.

OpenRouter API specifications
Context window
262,144 tokens
Maximum output
Provider-dependent
Inputs → output
Text → Text
Knowledge cutoff
Not verified

API model ID: nvidia/nemotron-3-ultra-550b-a55b

Capabilities & boundaries

What it supports. Where the limits are.

Tool support requires the appropriate API integration; a supported tool is not automatically active in every chat.

Supported API features and tools

  • NVIDIA’s large Nemotron reasoning model: 550B total parameters with 55B active, according to its official card
  • Text and code synthesis over long supplied inputs on the paid OpenRouter route
  • Configurable reasoning, with high and medium effort listed in OpenRouter’s model metadata
  • Tool requests and structured responses on compatible endpoints, including the reviewed DeepInfra route

Before you choose

  • Ultra’s parameter count and benchmark positioning do not establish a quality win on your workload. Compare it with Super using the same evidence and scoring rubric.
  • The NVIDIA card separates pre-training data through September 2025 from post-training through May 2026. There is no single unified knowledge-cutoff date claimed here, and neither date means live factual access.
  • The OpenRouter catalog reference is 262,144 tokens (256K). NVIDIA’s underlying model card describes up to 1M in suitable deployments; that is not the app’s confirmed route limit. Individual hosts can impose smaller windows. Leave room for the reply and reasoning.
  • This route accepts text and returns text, not image, audio or video input/output. Provide extracted text for document work; a prompt cannot add vision support.
  • Tool calls and structured outputs require a compatible host and integration. The model does not independently browse the web, execute code or access private files.
  • Answers, extracted fields and generated code can be wrong. Supply evidence and acceptance criteria, check cited passages and test code before use. A long context does not guarantee complete recall.

Choose the reasoning effort

mediumhigh · default

OpenRouter’s catalog lists reasoning enabled by default with high effort, and medium as another supported level. Whether these controls reach the selected host or appear in EZ Ai Assist depends on the integration. NVIDIA’s self-hosted enable_thinking chat-template flag is a different interface; do not assume it can be pasted into an OpenRouter request.

  • Use nvidia/nemotron-3-ultra-550b-a55b, without a :free suffix. Pin or record the actual provider when comparing behavior.
  • The reviewed DeepInfra endpoint lists 262,144 context and 16,384 maximum output tokens. Baseten instead lists 202,800 context and 182,520 output. These are endpoint ceilings, not a single universal Ultra output allowance.
  • Parameter support differs: DeepInfra advertises structured outputs, while the reviewed Baseten listing does not expose the same structured_outputs capability. Validate the exact schema and tool-choice mode before depending on it.

Put it to work

Start with a more useful prompt.

Original examples from EZ Ai Assist. Adapt these to your task and the features available in your workspace.

Workflow 01

Review an architectural migration

Turn a difficult design choice into verifiable decisions.

Compare these two migration proposals against the supplied requirements, failure history and operating constraints. Cite the relevant evidence for each conclusion. Identify assumptions, reversible steps, rollout checkpoints and rollback criteria. Recommend the smallest experiment that would distinguish the proposals. Do not invent benchmarks, costs or test results.

Workflow 02

Trace a subtle repository defect

Prioritize concrete failures over speculative refactoring.

Review these related source files and logs for the reported defect. Trace the relevant data flow, list competing explanations, and identify a minimal test that could rule each one in or out. Cite file names and functions. Suggest the smallest patch supported by the evidence, separating confirmed bugs from open questions. Do not claim you ran the tests.

Workflow 03

Reconcile conflicting reports

Build an evidence-backed account of a complex topic.

Compare the reports below using only the supplied material. Create a claim table with source, date, supporting passage, disagreement and confidence based on evidence quality. Distinguish differences in scope from direct contradictions. Finish with unresolved questions and what evidence would resolve them. Do not treat repeated claims as independent confirmation.

Developer reference

OpenRouter API pricing

These are named-provider OpenRouter API reference prices, not EZ Ai Assist subscription prices.

View EZ Ai Assist plans
DeepInfra route reference on OpenRouter · USD per 1,000,000 tokens
Token typePrice
Input$0.50
Output$2.20
Cached input$0.10
  • Rates were checked October 8, 2026. The named host is a pricing reference, not confirmation that the app pins that host. Routing, availability, cache eligibility and promotions can change the actual bill.
  • The DeepInfra reference is $0.50 input and $2.20 output per million tokens, with $0.10 cached input when eligible. Other hosts publish different rates.
  • For example, 100,000 uncached input tokens and 10,000 billed output tokens would be $0.072 at these reference rates, before any other charges. This is an illustration, not an app usage quote.
  • For a rough uncached estimate, multiply input tokens by the input rate and output tokens by the output rate, then divide by one million. Inspect the provider’s usage and billing records for reasoning and any additional charges.

Common questions

A few things worth knowing.

When should I try Ultra instead of Super?

Try Ultra for difficult code, multi-document analysis and complex planning, then score the answers against the same acceptance criteria. Keep Super if it meets your quality needs at lower measured cost; larger parameter counts do not decide the comparison.

Does Ultra have 256K or 1M context?

The paid OpenRouter catalog route is listed at 262,144 tokens. NVIDIA describes up to 1M for suitable underlying deployments, and some OpenRouter hosts offer less than 256K. Check the actual host and reserve space for reasoning and the final answer.

Why is there no single maximum output number?

Hosting limits differ substantially. The reviewed DeepInfra route lists 16,384 output tokens, while Baseten lists 182,520 with a smaller context window. Neither number establishes the app’s own limit or guarantees that much output for a full prompt.

Can I control how much reasoning it uses?

OpenRouter’s model metadata lists medium and high effort, defaulting to high. Host support and app controls still need to be checked. Ask for a concise justification and concrete verification steps; a longer response does not prove that a particular API setting was applied.

Are these the app’s confirmed model IDs?

Yes. The app owner confirmed these exact paid OpenRouter IDs, with text-only 256K routes and no :free variants. That confirmation does not test a live request or guarantee that every hosting feature and reasoning control is exposed in your workspace.

Are the token prices EZ Ai Assist subscription prices?

No. These are named-host developer reference rates on OpenRouter. Hosting choices, discounts and supported features can change; your EZ Ai Assist plan and any limits displayed in the app are separate.

Can I use this model to inspect an image or search the web?

Not from the model alone. These routes are text-only. Supply extracted text or use a verified vision model for images. Web search and other tools require an enabled integration; a generated tool request is not proof an action happened.

Check the source

Official documentation

Specifications and API prices checked on . Example prompts and workflow advice are editorial guidance from EZ Ai Assist.

Same provider

More from NVIDIA

NVIDIA

Nemotron 3.5 Super VL

NVIDIA’s vision-language reasoner for text, images and visual documents. The app uses paid OpenRouter hosting; context limits vary by serving host.

NVIDIA

Nemotron 3 Super

Balanced Nemotron workhorse for text, coding and tool-assisted workflows, using a paid 256K OpenRouter route.

NVIDIA

Nemotron 3.5 Lightning

Efficiency-focused Nemotron model for low-cost text workflows and rapid iteration, using a paid 256K OpenRouter route.

NVIDIA

Nemotron 3 Nano

Compact Nemotron model for short replies, extraction and bounded reasoning, using a paid 256K OpenRouter route.