Skip to content
aiqbench

NVIDIA · model passport

NVIDIA: Nemotron 3 Ultra

nemotron-3-ultra-550b-a55b
  • Open weights
  • 560.5B params
  • 262,144 context
  • Released 2026-06-03(HF repo created)
2/27axes covered
Best percentile
99.1IFBench (Artificial Analysis)
Cohort standing
59.2median 59.6

Thin coverage — read with caution

This model appears on only 2 of 27 benchmark axes considered. The overall standing has been shrunk toward the neutral 50 to reflect the small sample, and no weaknesses have been claimed — a low percentile on 2 axes is a coverage artifact, not a measured weakness.

What the data says

Benchmarked on 2 of 27 tracked axes. Every figure below is the model's own published result on that board.

On only 2 board(s), a ranked list of strengths would be false authority — the per-axis figures are shown below instead, without a verdict.

Uneven: 52.4 percentile points between its best and worst axis — 99.1 pct on IFBench (Artificial Analysis) versus 46.7 pct on LiveBench.

  • IFBench (Artificial Analysis)99.1 pct81.4%Verify →
  • LiveBench46.7 pct51.8%Verify →
  • OverallNot benchmarked
  • WebDev ArenaNot benchmarked
  • Vision ArenaNot benchmarked
  • Factuality ArenaNot benchmarked
  • Search ArenaNot benchmarked
  • Document ArenaNot benchmarked
  • Agent ArenaNot benchmarked
  • GPQA DiamondNot benchmarked
  • AIME (math)Not benchmarked
  • FrontierMathNot benchmarked
  • SimpleQA VerifiedNot benchmarked
  • ARC-AGI-2Not benchmarked
  • Humanity's Last ExamNot benchmarked
  • SimpleBenchNot benchmarked
  • BFCL v4 (function calling)Not benchmarked
  • SWE-bench Verified (live board)Not benchmarked
  • Terminal-Bench 4.0Not benchmarked
  • MMLU-Pro (Artificial Analysis)Not benchmarked
  • LiveCodeBench (Artificial Analysis)Not benchmarked
  • HHEM Hallucination (Vectara)Not benchmarked
  • SWE-bench VerifiedNot benchmarked
  • Terminal-BenchNot benchmarked
  • CodingNot benchmarked
  • Instruction followingNot benchmarked
  • Hard promptsNot benchmarked

Percentiles are the model's standing within each board's own population. Axes marked "not benchmarked" have no published result for this model — that is an absence of data, not a low score.

How to run it

Run it yourself

Repositorynvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16
Parameters560.5B
ArchitectureNemotronHForCausalLM

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

Ollama

We don't sync Ollama tags, so we won't print a command we can't verify — a wrong one would pull the wrong model. Search the Ollama library →

Unsloth fine-tuning recipes are model-specific and we don't derive them here. Unsloth docs →

Use it hosted

API model IDnvidia/nemotron-3-ultra-550b-a55b
Context length262,144 tokens
Max output182,520 tokens
Input price$0.60 per 1M tokens
Cached input price$0.12 per 1M tokens
Output price$2.40 per 1M tokens

API parameters this endpoint supports

Frequency penaltyInclude reasoningLogit biasMax tokensMin-pPresence penaltyReasoningReasoning effortRepetition penaltyResponse formatSeedStop sequencesStructured outputsTemperatureTool choiceTool callingTop-kTop-p

Example request

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","messages":[{"role":"user","content":"Hello"}]}'

View on OpenRouter →

Identifiers, limits and prices are synced from the upstream source shown above. Any field we haven't synced is left out entirely rather than guessed.

  • nemotron-3-ultra-550b-a55b stands at 59.2 (raw mean 72.9) — measured on only 2 of 27 axes, so the standing is shrunk toward the neutral 50.

    standingtemplateVerify →

  • nemotron-3-ultra-550b-a55b stands at 59.2 versus a cohort median of 59.6 across 211 models released within 180 days — below the peer median.

    cohorttemplateVerify →

Every claim is a computed fact from the current board data, phrased by a low-temperature model call and rejected by a validation gate if any number is not in the source row. How the leaderboards work →