// BENCHMARK

MMLU benchmark

AI model leaderboard for the MMLU benchmark. Compare how large language models score on MMLU, see the full ranking, and understand what this AI benchmark measures. Claude Sonnet 4 currently leads with 91.9%. 57-subject multiple-choice exam covering elementary to professional knowledge.

Leaderboard

# Model Organization Score Variant Source
#1 Claude Sonnet 4 Anthropic 91.9% official ↗
#2 gemini-2.5-pro Google DeepMind 91.9% official ↗
#3 Grok 4 xAI 91.9% official ↗
#4 GPT-4.1 OpenAI 90.2% official ↗
#5 gpt-oss-120b OpenAI 90% high reasoning official ↗
#6 gpt-oss-120b OpenAI 90% high official ↗
#7 GLM 4.5 Zhipu AI 90% official ↗
#8 LongCat-Flash Meituan Inc 89.71% official ↗
#9 DeepSeek-V3-0324 DeepSeek 89.5% official ↗
#10 Kimi K2 Moonshot 89.5% official ↗
#11 Pangu Pro MoE Huawei 89.3% instruct-em official ↗
#12 DeepSeek-R1-0528 DeepSeek 89.1% official ↗
#13 GPT-4 (Jun 2023) OpenAI 88.7% cited-nvlm official ↗
#14 Claude 3.5 Sonnet Anthropic 88.7% cited-nvlm official ↗
#15 Llama 3.1-405B Meta AI 88.6% instruct-cited-minimax-text official ↗
#16 DeepSeek-V3 DeepSeek 88.5% cited-minimax-text official ↗
#17 MiniMax-Text-01 MiniMax 88.5% from-hf-readme official ↗
#18 Claude 3.5 Sonnet Anthropic 88.3% 1022-cited-minimax-text official ↗
#19 Llama 3.1-405B Meta AI 88% instruct-cited-tulu3 official ↗
#20 DeepSeek-V3 DeepSeek 87.9% EM official ↗
#21 GPT-4o (Nov 2024) OpenAI 87.9% cited-tulu3 official ↗
#22 GLM-4.5-Air Zhipu AI 87.4% official ↗
#23 Llama 3.1-405B Meta AI 87.3% instruct-cited-nvlm official ↗
#24 Tulu 3 405B Allen Institute for AI 87% instruct-5-shot-cot official ↗
#25 Gemini 1.5 Pro Google DeepMind 86.8% 002-cited-minimax-text official ↗
#26 Gemini 2.0 Flash Google DeepMind 86.5% exp-cited-minimax-text official ↗
#27 Qwen2.5 Instruct (72B) Qwen 86.1% instruct-cited-minimax-text official ↗
#28 Llama 3.1-70B Meta AI 86% cited-telechat2 official ↗
#29 Gemini 1.5 Pro Google DeepMind 85.9% aug-2024-cited-nvlm official ↗
#30 GPT-4o (Nov 2024) OpenAI 85.7% cited-minimax-text official ↗
#31 Qwen2.5 Instruct (72B) Qwen 85.5% instruct-cited-tulu3 official ↗
#32 Llama 3.1-70B Meta AI 85.3% instruct-cited-tulu3 official ↗
#33 gpt-oss-20b OpenAI 85.3% high reasoning official ↗
#34 gpt-oss-20b OpenAI 85.3% high official ↗
#35 Hermes 3 405B Nous Research 84.9% cited-tulu3 official ↗
#36 Nemotron-H 56B NVIDIA 84.21% base-5-shot official ↗
#37 Llama 3.1-70B Meta AI 83.6% instruct-cited-nvlm official ↗
#38 Nemotron-H 47B NVIDIA 83.6% base-5-shot official ↗
#39 Qwen2-72B Qwen 82.3% instruct-cited-telechat2 official ↗
#40 Trinity-Large-Preview Arcee AI 82.3% official ↗
#41 Qwen-72B Qwen 82.3% instruct-cited-nvlm official ↗
#42 dots.llm1 Rednote 82.1% EM official ↗
#43 DeepSeek-V3 DeepSeek 82.1% cited-tulu3 official ↗
#44 NVLM-D 72B NVIDIA 82% megatron-text-from-hf-readme official ↗
#45 Llama 3-70B Meta AI 82% instruct-cited-nvlm official ↗
#46 TeleChat2-35B China Telecom 82% from-hf-readme official ↗
#47 Telechat2-115B China Telecom 80.9% from-hf-readme official ↗
#48 Mistral Small 3.1 Mistral AI 80.62% instruct-cited-msmall32 official ↗
#49 Mistral Small 3.2 Mistral AI 80.5% instruct-from-hf-readme official ↗
#50 Hermes 3 70B Nous Research 80.4% cited-tulu3 official ↗
#51 Qwen1.5-110B Qwen 80.4% cited-telechat2 official ↗
#52 TeleChat2-7B China Telecom 79.6% from-hf-readme official ↗
#53 Llama 4 Scout Meta AI 79.4% cited-pangu-pro-moe official ↗
#54 Phi-3.5-MoE Microsoft 78.9% instruct-5-shot official ↗
#55 Gemini 1.5 Flash (Sep 2024) Google DeepMind 78.7% cited-phi35moe official ↗
#56 Nemotron-4 340B NVIDIA 78.7% 0-shot-instruct-from-hf-readme official ↗
#57 Granite-4.0-H-Small IBM 78.44% 5-shot official ↗
#58 DeepSeek-V2 (MoE-236B) DeepSeek 77.8% cited-telechat2 official ↗
#59 OLMo 2 32B Allen Institute for AI 77.6% base-from-hf-readme official ↗
#60 GPT-4o mini OpenAI 77.2% cited-phi35moe official ↗
#61 Yi-1.5-34B 01.AI 77.1% cited-qwen2-57b official ↗
#62 Kimi Linear Moonshot 77% sft official ↗
#63 Qwen2-57B-A14B Qwen 76.5% base-from-hf-readme official ↗
#64 Nous-Hermes-2-Yi-34B Nous Research 75.5% cited-nvlm official ↗
#65 Gemma 2 9B Google DeepMind 74.6% instruct-cited-tulu3 official ↗
#66 Qwen2.5-7B Qwen 74.4% base-cited-olmo2 official ↗
#67 Qwen1.5-32B Qwen 74.3% cited-qwen2-57b official ↗
#68 IBM Granite 4.1 IBM 73.84%
#69 Qwen2.5-7B Qwen 73.5% instruct-5-shot-cited-falcon3 official ↗
#70 Aria Rhymes AI 73.3% 5-shot-from-hf-readme official ↗
#71 TeleChat2-3B China Telecom 72.9% from-hf-readme official ↗
#72 Nemotron-H 8B NVIDIA 72.77% base-5-shot official ↗
#73 Qwen2.5-7B Qwen 72.6% cited-phi4-mini official ↗
#74 Mixtral 8x7B Mistral AI 71.8% cited-qwen2-57b official ↗
#75 Llama 3.1-8B Meta AI 71.2% instruct-cited-tulu3 official ↗
#76 Gemma 2 9B Google DeepMind 70.6% base-cited-olmo2 official ↗
#77 Falcon3-7B Technology Innovation Institute 70.5% instruct-5-shot official ↗
#78 Qwen2-7B Qwen 70.3% base-from-hf-readme official ↗
#79 Llama 3.2 11B Meta AI 69.4% 5-shot-cited-aria official ↗
#80 Pixtral 12B Mistral AI 69.2% 5-shot-cited-aria official ↗
#81 Llama 3.1-8B Meta AI 69.15% instruct-cited-granite-3-2 official ↗
#82 Granite-4.0-H-Tiny IBM 68.65% 5-shot official ↗
#83 Llama 3.1-8B Meta AI 68.2% instruct-5-shot-cited-falcon3 official ↗
#84 Tulu 3 8B Allen Institute for AI 68.2% instruct-0-shot-cot official ↗
#85 Llama 3.1-8B Meta AI 68.1% instruct-cited-phi35moe official ↗
#86 Marin 8B Marin 67.6% base-5-shot-from-hf-readme official ↗
#87 Granite-4.0-H-Micro IBM 67.43% 5-shot official ↗
#88 Jamba AI21 Labs 67.4% cited-qwen2-57b official ↗
#89 Phi-4 Mini Microsoft 67.3% instruct-5-shot official ↗
#90 Mistral NeMo Mistral AI 67.2% instruct-cited-phi35moe official ↗
#91 Llama 3.1-8B Meta AI 66.9% base-cited-olmo2 official ↗
#92 Granite 3.2 8B IBM 66.79% instruct-from-hf-readme official ↗
#93 Granite 3.1 8B IBM 66.77% instruct-cited-granite-3-2 official ↗
#94 Llama 3-8B Meta AI 66.7% base-cited-falcon-mamba official ↗
#95 Llama 3.1-8B Meta AI 66.4% base-5-shot-cited-marin official ↗
#96 phi-3.5-mini Microsoft 65.5% cited-phi4-mini official ↗
#97 Qwen2.5-3B Qwen 65% cited-phi4-mini official ↗
#98 Gemma 7B Google DeepMind 64.6% base-cited-qwen2-7b official ↗
#99 Gemma 7B Google DeepMind 64.56% cited-falcon-mamba official ↗
#100 Mistral 7B Mistral AI 64.2% base-cited-qwen2-7b official ↗
#101 Mistral 7B Mistral AI 64.16% v0-1-cited-falcon-mamba official ↗
#102 DCLM 7B Apple 63.72% few-shot-from-hf-readme official ↗
#103 Falcon Mamba Technology Innovation Institute 62.11% base-from-hf-readme official ↗
#104 Llama 3.2 3B Meta AI 61.8% cited-phi4-mini official ↗
#105 Qwen1.5-7B Qwen 61% base-cited-qwen2-7b official ↗
#106 Ministral 8B Mistral AI 60.8% cited-phi4-mini official ↗
#107 DCLM 7B Apple 57.66% 0-shot-from-hf-readme official ↗
#108 Granite 3.2 2B IBM 57.18% instruct-from-hf-readme official ↗
#109 Granite 3.1 2B IBM 57.11% instruct-cited-granite-3-2 official ↗
#110 Qwen2-1.5B Qwen 56.5% base-from-hf-readme official ↗
#111 Llama 2-13B Meta AI 55.7% base-cited-olmo2 official ↗
#112 Phi-2 Microsoft 52.7% cited-qwen2 official ↗
#113 DeepSeek-R1-Distill-Qwen-7B DeepSeek 50.72% cited-granite-3-2 official ↗
#114 DeepSeek-R1-Distill-Qwen-7B Qwen 50.72% cited-granite-3-2 official ↗
#115 DeepSeek-R1-Distill-Llama-8B DeepSeek 45.8% cited-granite-3-2 official ↗
#116 Qwen2-0.5B Qwen 45.4% base-from-hf-readme official ↗
#117 Gemma 2B Google DeepMind 42.3% cited-qwen2 official ↗

Frequently asked questions about MMLU

What is the MMLU benchmark?

57-subject multiple-choice exam covering elementary to professional knowledge.

How is the MMLU benchmark scored?

MMLU 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?

As of the latest reported scores on GenAIList, Claude Sonnet 4 achieves the highest result on MMLU with a score of 91.9%.

Is a higher MMLU score better?

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