// BENCHMARK

MMLU-Redux benchmark

AI model leaderboard for the MMLU-Redux benchmark. Compare how large language models score on MMLU-Redux, see the full ranking, and understand what this AI benchmark measures. gemini-2.5-pro currently leads with 94.67%. Re-annotated, error-corrected version of MMLU.

Leaderboard

# Model Organization Score Variant Source
#1 gemini-2.5-pro Google DeepMind 94.67% lowthink-cited-ling-1t official ↗
#2 Kimi K2 Thinking Moonshot 94.4% thinking-no-tools official ↗
#3 DeepSeek-R1-0528 DeepSeek 93.4% reasoning official ↗
#4 GPT-5 OpenAI 92.75% main-cited-ling-1t official ↗
#5 Qwen3-235B-A22B Qwen 92.7% thinking official ↗
#6 Qwen3-235B-A22B-Thinking (Jul 2025) Qwen 92.7% reasoning official ↗
#7 Kimi K2 Moonshot 92.7% k2-0905-no-tools-cited-k2-thinking official ↗
#8 DeepSeek-V3.1-Terminus DeepSeek 92.37% non-thinking-cited-ling-1t official ↗
#9 EXAONE 4.0 (32B) LG AI Research 92.3% reasoning official ↗
#10 Llama 4 Maverick Meta AI 92.3% official ↗
#11 DeepSeek-V3-0324 DeepSeek 92.3% official ↗
#12 Ling-1T Ant Group 92.25% official ↗
#13 Kimi K2 Moonshot 91.58% instruct-0905-cited-ling-1t official ↗
#14 Qwen3-32B Qwen 90.9% reasoning official ↗
#15 Qwen3-32B Qwen 90.9% thinking official ↗
#16 Phi-4-Reasoning-plus Microsoft 90.8% reasoning official ↗
#17 Qwen3-30B-A3B Qwen 89.5% thinking official ↗
#18 Qwen3-235B-A22B Qwen 89.2% non-thinking official ↗
#19 DeepSeek-V3 DeepSeek 88.8% EM official ↗
#20 Qwen3-14B Qwen 88.6% thinking official ↗
#21 Qwen3-8B Qwen 87.5% thinking official ↗
#22 Qwen3-32B Qwen 85.7% non-thinking official ↗
#23 dots.llm1 Rednote 85.1% EM official ↗
#24 Gemma 3 27B Google DeepMind 85% official ↗
#25 Llama 4 Scout Meta AI 84.8% cited-pangu-pro-moe official ↗
#26 Qwen3-30B-A3B Qwen 84.1% non-thinking official ↗
#27 Qwen3-4B Qwen 83.7% thinking official ↗
#28 Qwen3-14B Qwen 82% non-thinking official ↗
#29 Pangu Pro MoE Huawei 81.5% instruct-em official ↗
#30 Kimi Linear Moonshot 80.3% sft official ↗
#31 Qwen3-8B Qwen 79.5% non-thinking official ↗
#32 Qwen3-4B Qwen 77.3% non-thinking official ↗
#33 Qwen3-1.7B Qwen 73.9% thinking official ↗
#34 EXAONE 4.0 (1.2B) LG AI Research 71.5% reasoning official ↗
#35 Qwen3-1.7B Qwen 64.4% non-thinking official ↗
#36 Qwen3-1.7B Qwen 63.4% non-reasoning official ↗
#37 Qwen3-0.6B Qwen 55.6% thinking official ↗
#38 Qwen3-0.6B Qwen 44.6% non-thinking official ↗
#39 Qwen3-0.6B Qwen 44.6% non-reasoning official ↗
#40 Gemma 3 1B Google DeepMind 40.9% non-reasoning official ↗

Frequently asked questions about MMLU-Redux

What is the MMLU-Redux benchmark?

Re-annotated, error-corrected version of MMLU.

How is the MMLU-Redux benchmark scored?

MMLU-Redux 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-Redux?

As of the latest reported scores on GenAIList, gemini-2.5-pro achieves the highest result on MMLU-Redux with a score of 94.67%.

Is a higher MMLU-Redux score better?

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