NVIDIA · Benchmarks

Nemotron 3 Ultra benchmarks

Nemotron 3 Ultra has 10 published scores in this dataset, from a NVIDIA release dated 2026-06-04. Its composite index is 74.4, which places it 18th of 22 ranked models and 13th among 15 open-weight models. Its strongest area is knowledge (index 99.6) and its weakest is agentic (index 50.8). Its best placing is on IFBench, where its 81.7 ranks 1st of 3 models with a published score.

Open weightsOfficial page

Category profile

Index 0–100: each score relative to the best published score on that test.

ReasoningCodingAgenticKnowledge
Reasoning
66.0#15 of 16 · 2 tests
Coding
81.4#8 of 13 · 3 tests
Agentic
50.8#8 of 8 · 2 tests
Knowledge
99.6#1 of 4 · 2 tests

Model facts

Lab
NVIDIA
Type
Open weights
Released
2026-06-04
Parameters
550B total, 55B active (MoE)
Context
1M
License
OpenMDW License 1.1

Nemotron 3 Ultra hardware requirements and cost

Even at 4-bit it needs about 379 GB (4× H200 (564 GB)), well beyond a home workstation, so most people run it on rented multi-GPU servers. Full BF16 precision needs about 1.3 TB (Multiple nodes (over 1.1 TB)).

Estimated GPU memory for Nemotron 3 Ultra
PrecisionWeightsWith headroomSmallest setup
BF16 (full precision)1.1 TB1.3 TBMultiple nodes (over 1.1 TB)
FP8 / 8-bit550 GB633 GB8× H100 (640 GB)
4-bit (GGUF Q4_K_M)330 GB379 GB4× H200 (564 GB)

Estimate: parameters × bytes per parameter, plus 15% for the KV cache and runtime at modest context lengths. Long contexts and large batches need more.

Lighter or stronger open models

Open models ranked just above and below on the composite index, with their 4-bit memory needs.

Nemotron 3 Ultra speed and pricing

Measured figures link to their source; estimates are labelled.

API output speed
161 tok/sArtificial Analysis median · Artificial Analysis
Time to first token
0.27 sOpenRouter model page provider stats
API price per 1M tokens
$0.500 in · $2.20 outcheapest provider on OpenRouter: DeepInfra · OpenRouter endpoints API
Compare speed and price across all models

Score by test

Rank is among every model with a published score on that test.

GPQA Diamond87.0%labelled 'GPQA (no tools)' on card#16 of 19 · 9.1 behind GPT-6 AstraHumanity's Last Exam26.7%no tools#19 of 23 · 37.7 behind Claude Opus 5.5Humanity's Last Exam (with tools)37.4%with tools#14 of 17 · 32.6 behind Kimi K3.1 ProLiveCodeBench89.0%LiveCodeBench v6#3 of 7 · 4.5 behind DeepSeek V4 ProSWE-bench Verified70.7%Real GitHub issues from popular Python repositories, fixed end to end.#4 of 5 · 10.9 behind Reflection Beam 1.0Terminal-Bench56.4%Terminal-Bench 2.1#13 of 13 · 34.2 behind DeepSeek V4.1 FlashBrowseComp44.4%Hard-to-find facts that need persistent web browsing.#8 of 8 · 47.7 behind Kimi K3.1 ProMMLU-Pro86.8%Broad academic knowledge with ten answer options instead of four.#2 of 5 · 0.7 behind DeepSeek V4 ProIFBench81.7%prompt loose#1 of 3 · top scoreLMArena Text Elo1427NVIDIA-Nemotron-3-Ultra-550B-A55B (NVFP4), ±7, board dated 2026-10-27#20 of 21 · 93 behind Gemini 4 Argon

Nearby in the composite ranking

Models with the closest composite index.

  1. 1DeepSeek V4.1 Flash79.8
  2. 2Gemma 4 31B79.2
  3. 3Step 5 Preview78.7
  4. 4Nemotron 3 Ultra74.4
  5. 5Gemma 4 26B A4B72.9
  6. 6Claude Haiku 5.572.4
  7. 7Gemini 3.8 Flash70.8

Head-to-head comparisons

Side-by-side pages built from the tests both models report.

No head-to-head page for Nemotron 3 Ultra yet. Compare it on the rankings dashboard.

Sources

ModelTestScoreSettingSource
Nemotron 3 UltraGPQA Diamond87.0labelled 'GPQA (no tools)' on cardHugging Face model card (NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16)
Nemotron 3 UltraHumanity's Last Exam26.7no toolsHugging Face model card (NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16)
Nemotron 3 UltraHumanity's Last Exam (with tools)37.4with toolsHugging Face model card (NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16)
Nemotron 3 UltraLiveCodeBench89.0LiveCodeBench v6Hugging Face model card (NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16)
Nemotron 3 UltraSWE-bench Verified70.7—Hugging Face model card (NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16)
Nemotron 3 UltraTerminal-Bench56.4Terminal-Bench 2.1Hugging Face model card (NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16)
Nemotron 3 UltraBrowseComp44.4—Hugging Face model card (NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16)
Nemotron 3 UltraMMLU-Pro86.8—Hugging Face model card (NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16)
Nemotron 3 UltraIFBench81.7prompt looseHugging Face model card (NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16)
Nemotron 3 UltraLMArena Text Elo1427NVIDIA-Nemotron-3-Ultra-550B-A55B (NVFP4), ±7, board dated 2026-10-27LMArena (arena.ai) text leaderboard

Scores are published results from model cards, lab reports, and public leaderboards, each linked to its source. Settings such as thinking mode or test version differ between sources and are shown next to each score. OpenLLM Buddy catalog models are uncensored fine-tunes; their scores are those of the official base model. Data collected 2026-10-30. How the index works