stealthmodels.

MODEL FIELD GUIDE

Nemotron 3 Ultra: Raccoon

Combines Mamba, attention and mixture-of-experts layers in a 550B-parameter model, with 55B active and multi-token prediction for faster serving.

Drawn by this model01 / 01
Raccoon benchmark result generated by this model
SVG BENCHMARKRaccoonView result

A good fit for

Self-hosted reasoning, tool use and retrieval over long documents on multi-GPU NVIDIA infrastructure.

Context window
1,000,000tokens[1]
Active parameters
55Bparameters[1]
License
OpenMDW License Agreement v1.1[1]
Model factsArchitecture, context & capabilities
First available
2026-06-04[1]
Developer
NVIDIA[1] [2]
Origin
United States[3]
Family
Nemotron 3 Ultra[1] [2]
Context window
1,000,000 tokens[1]
Architecture
LatentMoE: Mamba-2 + MoE + Attention hybrid with Multi-Token Prediction[1]
Total parameters
550B[1]
Active parameters per token
55B[1]
Input
text[1] [2]
Output
text[1] [2]
API access & pricingOpenRouter · $0.5 input / $2.2 output · per 1M tokens

OpenRouter via NVIDIA Build; input 0.5 USD/1M; output 2.2/1M; Partner endpoint price displayed by NVIDIA Build; not a universal model price.

[2]
  • OpenRouterObserved 2026-10-06
    Input $0.5Output $2.2Cached read $0.1

    per 1M tokens

    nvidia/nemotron-3-ultra-550b-a55b[4]
  • OpenRouter via NVIDIA BuildObserved 2026-10-06
    Input $0.5Output $2.2

    per 1M tokens

    Available

    nvidia/nemotron-3-ultra-550b-a55b[2]
  • DigitalOcean via NVIDIA BuildObserved 2026-10-06
    Input $0.9Output $1.7

    per 1M tokens

    Available

    nvidia/nemotron-3-ultra-550b-a55b[2]
Run it locallyWeights, memory & deployment

multi-GPU datacenter

[1] [6]
Measurements & sources6 primary references · Speed & serving conditions

StealthMark score

StealthMark measures fluid and visual intelligence of AI models.

General and Visual

General Intelligence

Perspective, anatomy, correct shadows, purposeful detail and physical consistency.

Visual Intelligence

Likeness, anatomy, detail, light, materials, composition and artistic expression.

Judging

Frontier Models blindly judge the AI generation unbiased through an intricate process. Our researchers control for mistakes.

Overall score

Overall combines weighted task scores on a scale from 0 to 100. Repeating a task does not increase its weight. Estimated means the total includes an estimate for one missing scene.

Benchmaxxing

Benchmaxx flags models that score unusually well on the familiar Pelican task compared with their other benchmark results. The percentage measures that imbalance, not the probability of training contamination. Flagged Pelican scores are excluded from Overall. Pelican never counts toward General or Visual.

We keep some prompts private to discourage test-specific optimization.

Scoring categories

Prompt & likeness

Subjects, actions and likeness.

Anatomy & construction

Coherent bodies, joints and machinery.

Space & placement

Perspective, scale and contact.

Light & reflections

Lighting, shadows and reflections.

Purposeful detail

Clear, purposeful small features.

Artistry & character

Composition, expression and drawing skill.

Materials & effects

Surfaces, texture and movement.

Visual integrity

Clean shapes, layers and edges.

BENCHMARK RESULTS

Compare results

This link opens the same results in the same order.

StealthMark benchmark results

Reset benchmark preferences?

Return to the default view, clear comparison selections and restore the advertisement.