// BENCHMARK

MMLU-Pro benchmark

AI model leaderboard for the MMLU-Pro benchmark. Compare how large language models score on MMLU-Pro, see the full ranking, and understand what this AI benchmark measures. Gemini 3 Pro currently leads with 90.1%. Harder MMLU successor with 10-option questions and reasoning-heavy items.

Leaderboard

# Model Organization Score Variant Source
#1 Gemini 3 Pro Google DeepMind 90.1% cited-deepseek-v3-2 official ↗
#2 Claude Opus 4.7 Anthropic 88.2% official ↗
#3 Claude Sonnet 4.5 Anthropic 88.2% cited-deepseek-v3-2 official ↗
#4 Intern-S2-Preview Shanghai AI Laboratory 88% official ↗
#5 GPT-5.5 OpenAI 87.5% no-tools official ↗
#6 GPT-5 OpenAI 87.5% cited-deepseek-v3-2 official ↗
#7 Claude Opus 4 Anthropic 87.3% official ↗
#8 NVIDIA Nemotron 3 Ultra NVIDIA 86.8% official ↗
#9 Grok 4 xAI 86.6% official ↗
#10 GPT-5 OpenAI 86.21% thinking-high-cited-ring-1t official ↗
#11 gemini-2.5-pro Google DeepMind 86.2% official ↗
#12 Gemini 3.1 Pro Google DeepMind 85.9% official ↗
#13 gemini-2.5-pro Google DeepMind 85.62% cited-ring-1t official ↗
#14 o3 OpenAI 85.3% official ↗
#15 DeepSeek-V3.2 DeepSeek 85% thinking official ↗
#16 MAI-Thinking-1 Microsoft 85% official ↗
#17 DeepSeek-R1-0528 DeepSeek 85% reasoning official ↗
#18 DeepSeek-V3.1-Terminus DeepSeek 85% thinking-cited-ring-1t official ↗
#19 DeepSeek-V3.2-Exp DeepSeek 85% from-hf-readme official ↗
#20 DeepSeek-R1-0528 DeepSeek 84.9% official ↗
#21 DeepSeek V4-Pro DeepSeek 84.7% official ↗
#22 Kimi K2 Thinking Moonshot 84.6% thinking-cited-deepseek-v3-2 official ↗
#23 GLM 4.5 Zhipu AI 84.6% official ↗
#24 Kimi K2 Thinking Moonshot 84.6% thinking-no-tools official ↗
#25 Qwen3-235B-A22B-Thinking (Jul 2025) Qwen 84.5% official ↗
#26 Qwen3-235B-A22B-Thinking (Jul 2025) Qwen 84.4% thinking-cited-ring-1t official ↗
#27 DeepSeek-V3.1-Terminus DeepSeek 83.25% non-thinking-cited-ling-1t official ↗
#28 Qwen3-235B-A22B-Thinking (Jul 2025) Qwen 83% reasoning official ↗
#29 LongCat-Flash Meituan Inc 82.68% official ↗
#30 Pangu Pro MoE Huawei 82.6% instruct-em official ↗
#31 Kimi K2.6 Moonshot 82.5% official ↗
#32 gemini-2.5-pro Google DeepMind 82.13% lowthink-cited-ling-1t official ↗
#33 Ling-1T Ant Group 82.04% official ↗
#34 MiniMax-M2 MiniMax 82% cited-deepseek-v3-2 official ↗
#35 GPT-5 OpenAI 81.94% main-cited-ling-1t official ↗
#36 Kimi K2 Moonshot 81.9% k2-0905-no-tools-cited-k2-thinking official ↗
#37 EXAONE 4.0 (32B) LG AI Research 81.8% reasoning official ↗
#38 GLM-5.1 Zhipu AI 81.4% official ↗
#39 GLM-4.5-Air Zhipu AI 81.4% official ↗
#40 DeepSeek-V3-0324 DeepSeek 81.2% official ↗
#41 MiniMax-M1-80k MiniMax 81.1% official ↗
#42 Kimi K2 Moonshot 81.03% instruct-0905-cited-ling-1t official ↗
#43 Trinity-Large-Thinking Arcee AI 80.7% official ↗
#44 MiniMax-M1-40k MiniMax 80.6% official ↗
#45 Ring-1T Ant Group 80.54% thinking official ↗
#46 Llama 4 Maverick Meta AI 80.5% official ↗
#47 Qwen3-32B Qwen 80% reasoning official ↗
#48 Llama Nemotron Super 49B NVIDIA 79.53% reasoning-on-cot-pass@1 official ↗
#49 DeepSeek V4-Flash DeepSeek 78.9% official ↗
#50 Muse Spark Meta AI 78.4% official ↗
#51 Claude 3.5 Sonnet Anthropic 78% 1022-cited-minimax-text official ↗
#52 Claude 3.5 Sonnet Anthropic 78% 1022-cited-mimo official ↗
#53 Gemma 4 12B Google DeepMind 77.2% official ↗
#54 Gemini 2.0 Flash Google DeepMind 76.4% exp-cited-minimax-text official ↗
#55 Phi-4-Reasoning-plus Microsoft 76% reasoning official ↗
#56 DeepSeek-V3 DeepSeek 76% EM official ↗
#57 DeepSeek-V3 DeepSeek 75.9% cited-minimax-text official ↗
#58 Gemini 1.5 Pro Google DeepMind 75.8% 002-cited-minimax-text official ↗
#59 MiniMax-Text-01 MiniMax 75.7% from-hf-readme official ↗
#60 GPT-4o (Nov 2024) OpenAI 74.4% cited-minimax-text official ↗
#61 Llama 4 Scout Meta AI 73.8% cited-pangu-pro-moe official ↗
#62 Llama 3.1-405B Meta AI 73.3% instruct-cited-minimax-text official ↗
#63 Gemma 4 31B Google DeepMind 72.5% official ↗
#64 Qwen2.5 Instruct (72B) Qwen 71.1% instruct-cited-minimax-text official ↗
#65 dots.llm1 Rednote 70.4% EM official ↗
#66 Gemma 4 26B Google DeepMind 70.1% official ↗
#67 Mistral Small 3.2 Mistral AI 69.06% instruct-5-shot-cot official ↗
#68 DeepSeek-R1-Distill-Qwen-14B DeepSeek 68.8% cited-mimo official ↗
#69 DeepSeek-R1-Distill-Qwen-14B Qwen 68.8% cited-mimo official ↗
#70 Gemma 3 27B Google DeepMind 67.5% official ↗
#71 Kimi Linear Moonshot 67.4% sft official ↗
#72 Mistral Small 3.1 Mistral AI 66.76% instruct-5-shot-cot official ↗
#73 GPT-4o mini OpenAI 62.8% cited-phi35moe official ↗
#74 Nemotron-H 47B NVIDIA 61.8% base-5-shot-cot official ↗
#75 Nemotron-H 56B NVIDIA 60.51% base-5-shot-cot official ↗
#76 EXAONE 4.0 (1.2B) LG AI Research 59.3% reasoning official ↗
#77 Gemini 1.5 Flash (Sep 2024) Google DeepMind 57.2% cited-phi35moe official ↗
#78 Qwen2.5-7B Qwen 56.2% cited-phi4-mini official ↗
#79 IBM Granite 4.1 IBM 55.99%
#80 Granite-4.0-H-Small IBM 55.47% 5-shot-cot official ↗
#81 Phi-3.5-MoE Microsoft 54.3% instruct-0-shot-cot official ↗
#82 DeepSeek-R1-Distill-Qwen-7B DeepSeek 53.5% cited-mimo official ↗
#83 DeepSeek-R1-Distill-Qwen-7B Qwen 53.5% cited-mimo official ↗
#84 Phi-4 Mini Microsoft 52.8% instruct-0-shot-cot official ↗
#85 Gemma 2 9B Google DeepMind 50.1% instruct-cited-phi35moe official ↗
#86 OLMo 2 32B Allen Institute for AI 49.7% base-from-hf-readme official ↗
#87 Yi-1.5-34B 01.AI 48.3% cited-qwen2-57b official ↗
#88 phi-3.5-mini Microsoft 47.4% cited-phi4-mini official ↗
#89 Qwen2.5-7B Qwen 45.8% base-cited-olmo2 official ↗
#90 Granite-4.0-H-Tiny IBM 44.94% 5-shot-cot official ↗
#91 Qwen2.5-3B Qwen 44.7% cited-phi4-mini official ↗
#92 Nemotron-H 8B NVIDIA 44.01% base-5-shot-cot official ↗
#93 Llama 3.1-8B Meta AI 44% instruct-cited-phi35moe official ↗
#94 Qwen1.5-32B Qwen 44% cited-qwen2-57b official ↗
#95 Qwen3-1.7B Qwen 43.7% non-reasoning official ↗
#96 Granite-4.0-H-Micro IBM 43.48% 5-shot-cot official ↗
#97 Qwen2.5-7B Qwen 43.1% instruct-5-shot-cited-falcon3 official ↗
#98 Qwen2-57B-A14B Qwen 43% base-from-hf-readme official ↗
#99 Gemma 2 9B Google DeepMind 42% base-cited-olmo2 official ↗
#100 Mixtral 8x7B Mistral AI 41% cited-qwen2-57b official ↗
#101 Falcon3-7B Technology Innovation Institute 40.7% instruct-5-shot official ↗
#102 Mistral NeMo Mistral AI 40.7% instruct-cited-phi35moe official ↗
#103 Qwen2-7B Qwen 40% base-from-hf-readme official ↗
#104 Llama 3.2 3B Meta AI 39.2% cited-phi4-mini official ↗
#105 Marin 8B Marin 36.5% base-from-hf-readme official ↗
#106 Llama 3.1-8B Meta AI 36.4% instruct-5-shot-cited-falcon3 official ↗
#107 Llama 3-8B Meta AI 35.4% base-cited-qwen2-7b official ↗
#108 Ministral 8B Mistral AI 35.3% cited-phi4-mini official ↗
#109 Llama 3.1-8B Meta AI 34.7% base-cited-olmo2 official ↗
#110 Gemma 7B Google DeepMind 33.7% base-cited-qwen2-7b official ↗
#111 Llama 3.1-8B Meta AI 33.3% base-cited-marin official ↗
#112 Mistral 7B Mistral AI 30.9% base-cited-qwen2-7b official ↗
#113 Qwen1.5-7B Qwen 29.9% base-cited-qwen2-7b official ↗
#114 Qwen3-0.6B Qwen 26.6% non-reasoning official ↗
#115 Llama 3.1-8B Meta AI 24.95% base-cited-falcon-mamba official ↗
#116 Llama 3-8B Meta AI 24.55% cited-falcon-mamba official ↗
#117 Llama 2-13B Meta AI 23.9% base-cited-olmo2 official ↗
#118 Mistral 7B Mistral AI 22.36% v0-1-cited-falcon-mamba official ↗
#119 Qwen2-1.5B Qwen 21.8% base-from-hf-readme official ↗
#120 Gemma 7B Google DeepMind 21.64% cited-falcon-mamba official ↗
#121 Gemma 2B Google DeepMind 15.9% cited-qwen2 official ↗
#122 Qwen2-0.5B Qwen 14.7% base-from-hf-readme official ↗
#123 Gemma 3 1B Google DeepMind 14.7% non-reasoning official ↗
#124 Falcon Mamba Technology Innovation Institute 14.47% 5-shot-from-hf-readme official ↗

Frequently asked questions about MMLU-Pro

What is the MMLU-Pro benchmark?

Harder MMLU successor with 10-option questions and reasoning-heavy items.

How is the MMLU-Pro benchmark scored?

MMLU-Pro is scored using the accuracy (%) metric, where a higher score is better. The maximum achievable score is 100.000. GenAIList aggregates reported scores from model providers and papers into a single ranked leaderboard.

Which AI model scores highest on MMLU-Pro?

As of the latest reported scores on GenAIList, Gemini 3 Pro achieves the highest result on MMLU-Pro with a score of 90.1%.

Is a higher MMLU-Pro score better?

Yes. On MMLU-Pro a higher score indicates better performance, so models near the top of the leaderboard are the strongest.