# Nemotron 3 Ultra

The largest open model NVIDIA has released. Nemotron 3 Ultra activates 55B of its 550B parameters per token, holds a context window of 1M tokens, and runs the reasoning layer of long-horizon agent systems.

- **Model ID:** `nvidia/nemotron-3-ultra-550b-a55b`
- **Type:** chat
- **Providers:** togetherai, deepinfra, baseten
- **Context window:** 1,000,000
- **Maximum output tokens:** 65,000
- **Pricing:** $0.5/1M input tokens, $2.4/1M output tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/nemotron-3-ultra-550b-a55b

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { streamText } from 'ai'

const result = streamText({
  model: 'nvidia/nemotron-3-ultra-550b-a55b',
  prompt: 'Why is the sky blue?'
})
```

## About

NVIDIA released Nemotron 3 Ultra on June 4, 2026 as the largest model in the Nemotron 3 family, completing the tier above Nano and Super. It carries 550B total parameters with 55B active per token, and NVIDIA positions it as the reasoning and orchestration layer for long-running agent workflows: the model that handles planning, synthesis, and verification while lighter models execute routine steps.

The architecture interleaves three layer types. Mamba layers process long sequences with linear-time complexity, which keeps a context window of 1M tokens practical. Transformer attention layers appear at select depths to preserve precise recall from large contexts. Latent mixture-of-experts (MoE) routing compresses token embeddings into a smaller latent space before selecting experts, so distinct specialists activate for reasoning, coding, and tool calls without dense compute. Multi-token prediction (MTP) layers predict several future tokens per forward pass, providing built-in speculative decoding for long outputs.

Nemotron 3 Ultra scores 91% on PinchBench, 82% on IFBench, and 95% on Ruler at 1M tokens. Weights, data, and recipes are released under the Linux Foundation's permissive OpenMDW-1.1 license. Full details: https://www.together.ai/models/nvidia-nemotron-3-ultra.

## What to consider

Long-running agent sessions accumulate tokens quickly, and a context window of 1M tokens makes it easy to carry everything forward. Budget for that before you scale. Compare $0.5 and $2.4, and use prompt caching at $0.12 for repeated prefixes like system prompts and tool definitions.

Output is capped at 65K tokens per request, so plan chunking for very long generations. Nemotron 3 Ultra is the flagship tier of the Nemotron 3 family. Reserve it for the planning and verification calls that need the depth, and route routine steps to smaller Nemotron 3 models.

## When to use

### Best For

- Agent orchestration: planning, synthesis, and verification steps in long-running multi-agent pipelines
- Long-horizon coding agents working across large codebases and extended tool-call sequences
- Deep research workflows that gather, cross-check, and synthesize evidence in a single context
- Full-context tasks that keep complete agent histories, codebases, or document sets in one pass
- Teams that need open weights and a permissive license for governance or reproducibility

### Consider Alternatives When

- Routine pipeline steps run fine on Nemotron 3 Nano at far lower compute
- Nemotron 3 Super covers your planning depth at a smaller footprint
- You need vision or multimodal inputs. Nemotron 3 Ultra is a text reasoning model
- Cost per token is the primary constraint and a lighter model meets your quality bar

## Best for

- **Agent Orchestration Backbones:** Planning, synthesis, and verification steps in long-running multi-agent pipelines
- **Long-Horizon Coding Agents:** Multi-step software tasks that span large codebases and extended tool-call sequences
- **Deep Research Workflows:** Gathering, cross-checking, and synthesizing evidence across many sources in one context
- **Full-Context Session Handling:** Keeping complete agent histories, codebases, or document sets in a single pass
- **Open-Model Requirements:** Teams that need open weights and permissive licensing for governance or reproducibility

## Consider alternatives

- **Lightweight Task Execution:** Nemotron 3 Nano handles routine pipeline steps at far lower compute
- **Mid-Tier Agent Planning:** Nemotron 3 Super covers complex multi-agent decisions at a smaller footprint
- **Vision or Multimodal Inputs:** Nemotron 3 Ultra is a text reasoning model, so image and video tasks need a vision-language model
- **Cost-First Workloads:** A smaller model may deliver acceptable quality at lower per-token rates

## Frequently asked questions

### Why does "550B total, 55B active" matter for inference cost?

You pay compute proportional to the active parameters, not the total. Nemotron 3 Ultra draws on 550B parameters of learned knowledge while activating 55B per token. Latent MoE routing selects the relevant experts in a compressed latent space, which keeps routing overhead low.

### How does Nemotron 3 Ultra fit alongside Nemotron 3 Nano and Super?

Nemotron 3 Ultra is the largest tier of the Nemotron 3 family. Nano (30B/3B active) handles high-throughput lightweight steps, Super (120B/12B active) covers mid-tier multi-agent planning, and Nemotron 3 Ultra takes the hardest planning, synthesis, and verification calls. Many pipelines pair a smaller tier for routine steps with Nemotron 3 Ultra for complex decisions.

### How does Nemotron 3 Ultra handle a context window of 1M tokens?

Mamba layers process sequences with linear-time complexity, so long contexts avoid the quadratic memory growth of pure attention. Nemotron 3 Ultra scores 95% on the Ruler benchmark at 1M tokens, which tests recall across very long inputs.

### Is Nemotron 3 Ultra open source?

Yes. NVIDIA released the weights, data, and recipes under the Linux Foundation's OpenMDW-1.1 license, a permissive license built for open model distributions.

### Does Nemotron 3 Ultra support Zero Data Retention?

Yes, Zero Data Retention is available for this model. Zero Data Retention is offered on a per-provider basis. See https://vercel.com/docs/ai-gateway/capabilities/zdr for details.

### Where are per-token prices listed?

Rates are listed on this page. They reflect the providers routing through AI Gateway and shift when providers update their pricing.

## Links

- [Model page](https://vercel.com/ai-gateway/models/nemotron-3-ultra-550b-a55b)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://www.together.ai/models/nvidia-nemotron-3-ultra)