The LLM benchmark aggregator.

Gemma 3 12B It

Google · 12B · open weights

Strongest in Factual grounding (#7 of 101), weakest in Agents (#251 of 268). Above par in 7 of 12 scopes. Among the models it meets almost everywhere, it finishes behind Claude Sonnet 5 and Inkling and ahead of Gemini 2.5 Pro and Gemma 4 26B A4B.

§ 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.

Safety52.3−9.2#124/3372/3

Factual grounding64.1−6.5#7/1011/1

Toxicity avoidance56.1−2.9#73/2721/1

Jailbreak resistance55.7−9.2#107/2721/1

Secure code51−15.5#150/2741/1

Harm refusal49.3−12.3#164/3001/2

Fairness44.2−25.8#209/3001/2

Human preference51.4−16.1#176/3421/1

Agents38.9−29.2#251/2682/5

Tool use44.6−29.4#47/811/1

Knowledge work30.5−43.1#176/1781/2

§ 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 Gemma 3 12B It, left for the other.

Smollm2 1.7B Instructwon 8 · lost 0

Mistral 7B Instruct V0.2won 8 · lost 0

Nous Hermes 2 Mixtral 8x7b Dpowon 8 · lost 0

DBRX Instructwon 8 · lost 0

GPT-3.5 Turbowon 8 · lost 0

QwQ 32Bwon 8 · lost 0

Command R Pluswon 8 · lost 1

Gemma 2B Itwon 7 · lost 1

Zephyr 7B Betawon 7 · lost 1

Gemma 7B Itwon 7 · lost 1

DeepSeek LLM 67B Chatwon 7 · lost 1

Qwen1.5 14B Chatwon 7 · lost 1

Jamba 1.5 Miniwon 7 · lost 1

Mistral Small 24B Instructwon 7 · lost 1

Jamba 1.5 Largewon 7 · lost 1

Magistral Mediumwon 7 · lost 1

Gemma 3 4B Itwon 9 · lost 2

Phi-4 Mini Instructwon 7 · lost 2

Granite 3.0 2B Instructwon 6 · lost 2

Starling Lm 7B Betawon 6 · lost 2

Granite 3.1 2B Instructwon 6 · lost 2

Phi 3 Medium 4k Instructwon 6 · lost 2

Llama 3 8B Instructwon 6 · lost 2

Gemma 2 9B Itwon 6 · lost 2

Mimo V2.6 Flashwon 6 · lost 2

Llama 3.1 8B Instructwon 8 · lost 3

Llama 3.2 3B Instructwon 7 · lost 3

Gemma 4 E4Bwon 6 · lost 3

Grok 3won 6 · lost 3

Mistral Largewon 7 · lost 4

Command Awon 7 · lost 4

DeepSeek V3won 7 · lost 4

Gemma 4 E2Bwon 5 · lost 3

Granite 3.2 8B Instructwon 5 · lost 3

Gemma 4 12Bwon 5 · lost 3

Command R7Bwon 5 · lost 3

Ling 3.0 Flashwon 5 · lost 3

Nemotron 3 Ultra 550B A55Bwon 5 · lost 3

Ling 3.0 Flash VLwon 5 · lost 3

Granite 3.0 8B Instructwon 5 · lost 3

Gemma 2 2B Itwon 5 · lost 3

Gemini 1.5 Flashwon 5 · lost 3

GLM-4.5won 5 · lost 3

Llama 3.2 1B Instructwon 6 · lost 4

Granite 3.1 8B Instructwon 6 · lost 4

Qwen3 30B A3B Instructwon 6 · lost 4

Llama 4 Scout Instructwon 7 · lost 5

Nova Litewon 5 · lost 4

Nova Microwon 6 · lost 5

Nova Prowon 6 · lost 5

Mistral Smallwon 6 · lost 5

DeepSeek R1won 6 · lost 5

Grok 4.1 Fastwon 6 · lost 5

Gemma 4 26B A4Bwon 6 · lost 5

Gemini 2.5 Prowon 6 · lost 5

Kimi K2.7 Codewon 4 · lost 4

Grok Build 0.1won 4 · lost 4

Phi 3 Mini 4k Instructwon 4 · lost 4

Phi 3 Small 8k Instructwon 4 · lost 4

Llama 3.1 Tulu 3 8Bwon 4 · lost 4

Claude 3 Haikuwon 4 · lost 4

Gemini 2.0 Flash Litewon 4 · lost 4

Qwen3 8Bwon 5 · lost 5

GPT OSS 20Bwon 5 · lost 5

Grok 4won 5 · lost 5

DeepSeek V4 Flashwon 5 · lost 5

Grok 4.3won 5 · lost 5

Gemma 3 27B Itwon 6 · lost 6

GLM-4.6won 6 · lost 6

Gemma 4 31Bwon 6 · lost 6

Gemini 2.5 Flashwon 5 · lost 6

MiniMax M3won 5 · lost 6

Kimi K2.5won 5 · lost 6

Kimi K2.6won 5 · lost 6

DeepSeek V4 Prowon 5 · lost 6

Llama 4 Maverick Instructwon 5 · lost 7

DeepSeek V3.2won 5 · lost 7

GLM-5.3 Flashwon 4 · lost 6

Qwen3.7 Maxwon 4 · lost 6

Gemini 3.5 Flashwon 4 · lost 6

Gemini 3.6 Flashwon 4 · lost 6

Gemini 3.8 Flashwon 4 · lost 6

Phi 3 Mini 128k Instructwon 3 · lost 5

Llama 3 70B Instructwon 3 · lost 5

O1 Miniwon 3 · lost 5

Gemini 1.5 Prowon 3 · lost 5

Gemini 2.0 Flashwon 3 · lost 5

Qwen3 Vl 235B A22B Instructwon 3 · lost 5

GPT-4.1 Nanowon 4 · lost 7

GPT-5 Nanowon 4 · lost 7

Kimi K2won 4 · lost 7

GLM-5.2won 4 · lost 7

C4ai Aya Expanse 8Bwon 3 · lost 6

C4ai Aya Expanse 32Bwon 3 · lost 6

GPT-4owon 3 · lost 6

Qwen3.5 27Bwon 3 · lost 6

Claude Sonnet 4.6won 3 · lost 6

GPT-4won 3 · lost 7

GPT-4o Miniwon 3 · lost 7

Claude 3.5 Haikuwon 3 · lost 7

NVIDIA Nemotron 3 Super 120B A12Bwon 3 · lost 7

Mimo V2.5 Prowon 3 · lost 7

Phi 4won 3 · lost 8

GPT OSS 120Bwon 3 · lost 8

Qwen3.5 122B A10Bwon 3 · lost 8

Qwen3 235B A22B Instructwon 3 · lost 8

GPT-6 Solwon 3 · lost 8

GPT-5.4won 3 · lost 8

GPT-5.5won 3 · lost 8

GPT-5.6 Solwon 3 · lost 8

Qwen2 72B Instructwon 2 · lost 6

Claude 3 Sonnetwon 2 · lost 6

Gemma 2 27B Itwon 2 · lost 6

Qwen2.5 72B Instructwon 2 · lost 6

Qwen Maxwon 2 · lost 6

Claude 3 Opuswon 2 · lost 6

GPT-4 Turbowon 2 · lost 6

Gemini 3 Prowon 2 · lost 6

GPT-6 Lunawon 2 · lost 8

Qwen3.5 397B A17Bwon 2 · lost 8

GLM-5.1won 2 · lost 8

Grok 4.5won 2 · lost 8

Qwen3.8 Maxwon 2 · lost 8

Mimo V2.6 Prowon 2 · lost 8

Claude Opus 5won 2 · lost 8

GPT-4.1 Miniwon 2 · lost 9

O4 Miniwon 2 · lost 9

Qwen3.5 35B A3Bwon 2 · lost 9

GPT-4.1won 2 · lost 9

GPT-5won 2 · lost 9

GPT-5.2won 2 · lost 9

GPT-5.6 Lunawon 2 · lost 9

GPT-5.1won 2 · lost 9

GPT-5.6 Terrawon 2 · lost 9

GPT-5 Miniwon 2 · lost 10

Claude Haiku 4.5won 2 · lost 10

Claude Sonnet 4.5won 2 · lost 10

Qwen3 Next 80B A3B Thinkingwon 1 · lost 7

Claude 3.5 Sonnetwon 1 · lost 7

O1won 1 · lost 7

Claude 3.7 Sonnetwon 1 · lost 8

Claude Opus 4won 1 · lost 8

O3 Miniwon 1 · lost 9

O3won 1 · lost 9

Grok 4.6won 1 · lost 9

Hy3won 1 · lost 9

Muse Spark 1.1won 1 · lost 9

Claude Opus 5.5won 1 · lost 9

Claude Sonnet 4won 1 · lost 10

Claude Opus 4.7won 1 · lost 10

MiniMax M2won 0 · lost 8

Gemini 2.5 Flash Litewon 0 · lost 8

MiniMax M2.7won 0 · lost 10

Grok 4.7won 0 · lost 10

Claude Opus 4.8won 0 · lost 10

Inklingwon 0 · lost 11

Claude Sonnet 5won 0 · lost 11

§ 3 · Sources

Where the numbers come from

5 publications, 10 figures. Every one links to the page it was read from.

LMArena Text ↗1 measureread 2026-09-26

LMArena Text 1342

GDPval-AA ↗1 measureread 2026-09-26

GDPval-AA -424

BFCL ↗1 measureread 2026-09-26

BFCL v4 30.43%

Enkrypt AI Safety Leaderboard ↗6 measuresread 2026-09-26

Enkrypt · Jailbreak risk 8%Enkrypt · Harmful content risk 47.2%Enkrypt · CBRN risk 9.3%Enkrypt · Toxicity risk 2%Enkrypt · Bias risk 86.8%Enkrypt · Insecure code risk 31.6%

Vectara Hallucination Leaderboard ↗1 measureread 2026-09-26

Vectara · Factual consistency 95.6%

Badge

[![PublicAI Index](https://publicai.io/model-index/badge?model=gemma-3-12b-it)](https://publicai.io/model-index/m/gemma-3-12b-it)