The LLM benchmark aggregator.

Nemotron 3 Ultra

NVIDIA · 561B · open weights

Strongest in Science (#44 of 122), weakest in Coding (#145 of 165). Above par in 8 of 22 scopes. Among the models it meets almost everywhere, it finishes behind Claude Sonnet 5 and Claude Opus 5 and ahead of Gemma 4 31B and Llama 4 Maverick Instruct.

§ 1 · Profile

What it is good at

Bars run from 50 — the average of the models each source lists — so right of the line is above par. The middle column is the gap to whoever leads that scope.

Knowledge54.7−8.8#59/1381/2

Academic knowledge55.7−8.2#54/1231/1

Factuality52.6−6.4—/0✱✱

Reasoning50.7−16.2#86/1782/4

Science57−6.5#44/1221/1

Reasoning46.2−22.9#75/1401/3

Expert reasoning49.6−11.3—/0✱✱

Professional48.1−13.5#104/1681/1

Finance51.2−11.7#80/1511/1

Medical46.7−20.1#100/1401/1

Legal45.2−21.9#102/1511/1

Core abilities48.9−18.5#109/2041/3

General intelligence48.3−21.8#112/2041/3

Long context53.2−11.4—/0✱✱

Coding41.6−27.5#145/1651/5

Agentic coding40.6−28.1#142/1571/4

Code generation52−18—/77✱0/2

Agents40.3−27.8—/268✱0/5

Knowledge work33.6−40—/178✱0/2

Tool use34.2−39.8—/81✱0/1

Web research31.8−32.4—/4✱0/1

Workflow automation35.6−26.4—/0✱✱

§ 2 · Head to head

What it beats, and what beats it

The same models turn up scope after scope. Counted once: where both were placed, who finished higher. Bars run right for Nemotron 3 Ultra, left for the other.

GPT-4o Miniwon 13 · lost 2

Mistral Medium 3.5won 12 · lost 2

Llama 4 Scout Instructwon 11 · lost 3

Command Awon 10 · lost 4

GPT-5.4 Nanowon 11 · lost 5

GPT-4.1 Nanowon 11 · lost 5

GPT-5.4 Miniwon 10 · lost 5

Mistral Large 3won 10 · lost 5

O3 Miniwon 9 · lost 5

Qwen3.6 35B A3Bwon 9 · lost 5

Claude Haiku 4.5won 10 · lost 6

GPT OSS 120Bwon 9 · lost 6

Grok 4.3won 9 · lost 7

Llama 4 Maverick Instructwon 9 · lost 7

Gemma 4 31Bwon 9 · lost 8

K2 Horizon MoVA 36B A4Bwon 7 · lost 7

Mimo V2.5 Prowon 7 · lost 7

Gemini 3.1 Flash Litewon 7 · lost 7

DeepSeek V3won 8 · lost 8

Claude Sonnet 4won 8 · lost 8

DeepSeek R1won 7 · lost 8

Qwen3.6 Pluswon 7 · lost 9

GPT-4.1 Miniwon 7 · lost 9

GLM-4.6won 6 · lost 9

Kimi K2won 6 · lost 9

Qwen3 235B A22B Instructwon 6 · lost 9

Gemini 3.5 Flash Litewon 6 · lost 10

O4 Miniwon 6 · lost 10

MiniMax M3won 8 · lost 14

GPT-4.1won 5 · lost 10

DeepSeek V3.2won 5 · lost 11

Gemini 2.5 Flashwon 5 · lost 11

Qwen3.8 27Bwon 6 · lost 14

Claude Opus 4won 4 · lost 10

Kimi K2.6won 4 · lost 12

GPT-5 Miniwon 4 · lost 12

Grok 4won 3 · lost 12

GLM-5.3 Flashwon 3 · lost 13

Claude Sonnet 4.6won 3 · lost 13

O3won 3 · lost 13

Gemini 2.5 Prowon 3 · lost 13

Inklingwon 4 · lost 18

GLM-5.2won 3 · lost 18

GPT-5.6 Lunawon 3 · lost 19

Kimi K2.5won 2 · lost 13

Muse Spark 1.1won 2 · lost 14

Gemini 3.6 Flashwon 2 · lost 14

Qwen3.7 Maxwon 2 · lost 14

K2 Horizon 375B A23Bwon 2 · lost 14

DeepSeek V4.1 Flashwon 2 · lost 15

DeepSeek V4 Flashwon 2 · lost 16

DeepSeek V4 Prowon 2 · lost 16

GPT-5.6 Terrawon 2 · lost 19

Claude Opus 4.6won 1 · lost 13

Muse Spark 1.2won 1 · lost 14

GPT-5won 1 · lost 14

Claude Sonnet 4.5won 1 · lost 15

Claude Opus 4.5won 1 · lost 15

GPT-5.2won 1 · lost 15

Gemini 3.8 Flashwon 1 · lost 15

GPT-5.4won 1 · lost 15

Grok 4.6won 1 · lost 15

Qwen3.8 Maxwon 1 · lost 15

Gemini 3.1 Prowon 1 · lost 15

Gemini 3.5 Flashwon 1 · lost 16

Grok 4.5won 1 · lost 16

Kimi K3won 1 · lost 17

GPT-5.6 Solwon 1 · lost 18

Gemini 3 Prowon 0 · lost 15

Claude Opus 4.8won 0 · lost 16

Claude Opus 4.7won 0 · lost 16

Gemini 3.7 Flashwon 0 · lost 16

GPT-5.5won 0 · lost 16

Claude Fable 5.1won 0 · lost 16

Claude Fable 5won 0 · lost 17

GLM-5.3won 0 · lost 18

Claude Opus 5won 0 · lost 18

Claude Sonnet 5won 0 · lost 22

§ 3 · Sources

Where the numbers come from

5 publications, 36 figures. Every one links to the page it was read from.

Artificial Analysis ↗1 measureread 2026-09-26

Artificial Analysis Intelligence Index 23

SimpleBench ↗1 measureread 2026-09-26

SimpleBench 41.7%

Vals.ai ↗12 measuresread 2026-09-26

Vals · Legal Research Bench 15.38%Vals · LegalBench 82.07%Vals · Harvey Legal Agent Benchmark 0.42%Vals · Finance Agent 37.67%Vals · CorpFin 65.46%Vals · TaxEval 73.1%Vals · MedCode 38.62%Vals · SWE-bench Verified 69% (Mini-SWE-agent)Vals · Vibe Code Bench 7.64% (OpenHands)Vals · Code Migration 4.91%Vals · GPQA Diamond 86.11%Vals · MMLU Pro 85.76%

✱✱ IFM — K2 Horizon launch ↗the model’s own publisher17 measuresread 2026-09-26

GDPVal-AA 1162tau3-Banking 14.2%Toolathlon Verified 34.3%Automation Bench Public 8%Apex-Agents (pass@1) 9%MCPMark 45.7%BrowseComp 44.4%WildClawBench 34.2%Terminal-Bench 2.1 53.9%SciCode 39.9%SWE Bench Pro 38.7%Humanity's Last Exam (without tools) 28.4%GPQA Diamond 86.7%CritPt 3.1%AA-LCR 71%AA-Omniscience Accuracy 23%AA-Omniscience Non-Hallucination 70%

✱ CellCog — K2 Horizon analysis ↗5 measuresread 2026-09-26

GDPVal-AA (Elo) 1162Toolathlon Verified 34.3%Terminal-Bench 2.1 53.9%SWE-bench Pro (strict) 38.7%MCPMark 45.7%

Badge

[![PublicAI Index](https://publicai.io/model-index/badge?model=nemotron-3-ultra)](https://publicai.io/model-index/m/nemotron-3-ultra)