New: connect Claude & other AIs to GenAIList over MCP — research the catalog and contribute to the shared knowledge base. Learn how →
MMMLU benchmark
AI model leaderboard for the MMMLU benchmark. Compare how large language models score on MMMLU, see the full ranking, and understand what this AI benchmark measures. Claude Opus 4.5 currently leads with 90.8%. Multilingual MMLU professionally translated into 14 languages.
Leaderboard
| # | Model | Organization | Score | Variant | Source |
|---|---|---|---|---|---|
| #1 | Claude Opus 4.5 | Anthropic | 90.8% | official ↗ | |
| #2 | Qwen3-235B-A22B | Qwen | 84.3% | thinking-14lang | official ↗ |
| #3 | gpt-oss-120b | OpenAI | 81.3% | high reasoning | official ↗ |
| #4 | gpt-oss-120b | OpenAI | 81.3% | high-avg-14lang | official ↗ |
| #5 | Qwen3-32B | Qwen | 80.6% | thinking-14lang | official ↗ |
| #6 | Qwen3-235B-A22B | Qwen | 79.8% | non-thinking-14lang | official ↗ |
| #7 | Qwen3-30B-A3B | Qwen | 78.4% | thinking-14lang | official ↗ |
| #8 | Qwen3-14B | Qwen | 77.9% | thinking-14lang | official ↗ |
| #9 | Gemini 1.5 Flash (Sep 2024) | Google DeepMind | 77.2% | cited-phi35moe | official ↗ |
| #10 | Qwen3-32B | Qwen | 76.5% | non-thinking-14lang | official ↗ |
| #11 | gpt-oss-20b | OpenAI | 75.7% | high reasoning | official ↗ |
| #12 | gpt-oss-20b | OpenAI | 75.7% | high-avg-14lang | official ↗ |
| #13 | Qwen3-8B | Qwen | 74.4% | thinking-14lang | official ↗ |
| #14 | Qwen3-30B-A3B | Qwen | 73.8% | non-thinking-14lang | official ↗ |
| #15 | GPT-4o mini | OpenAI | 72.9% | cited-phi35moe | official ↗ |
| #16 | Qwen3-14B | Qwen | 72.6% | non-thinking-14lang | official ↗ |
| #17 | Phi-3.5-MoE | Microsoft | 69.9% | instruct-multilingual-5-shot | official ↗ |
| #18 | Qwen3-4B | Qwen | 69.8% | thinking-14lang | official ↗ |
| #19 | Granite-4.0-H-Small | IBM | 69.69% | 5-shot | official ↗ |
| #20 | Qwen3-8B | Qwen | 66.9% | non-thinking-14lang | official ↗ |
| #21 | Qwen2.5-7B | Qwen | 64.4% | cited-phi4-mini | official ↗ |
| #22 | Gemma 2 9B | Google DeepMind | 63.8% | instruct-cited-phi35moe | official ↗ |
| #23 | Granite-4.0-H-Tiny | IBM | 61.87% | 5-shot | official ↗ |
| #24 | Qwen3-4B | Qwen | 61.7% | non-thinking-14lang | official ↗ |
| #25 | Qwen3-1.7B | Qwen | 59.1% | thinking-14lang | official ↗ |
| #26 | Mistral NeMo | Mistral AI | 58.9% | instruct-cited-phi35moe | official ↗ |
| #27 | Llama 3.1-8B | Meta AI | 56.2% | instruct-cited-phi35moe | official ↗ |
| #28 | Qwen2.5-3B | Qwen | 55.9% | cited-phi4-mini | official ↗ |
| #29 | Granite-4.0-H-Micro | IBM | 55.19% | 5-shot | official ↗ |
| #30 | phi-3.5-mini | Microsoft | 51.8% | cited-phi4-mini | official ↗ |
| #31 | Phi-4 Mini | Microsoft | 49.3% | instruct-5-shot | official ↗ |
| #32 | Qwen3-1.7B | Qwen | 48.3% | non-thinking-14lang | official ↗ |
| #33 | Llama 3.2 3B | Meta AI | 48.1% | cited-phi4-mini | official ↗ |
| #34 | Ministral 8B | Mistral AI | 46.4% | cited-phi4-mini | official ↗ |
| #35 | Qwen3-0.6B | Qwen | 43.1% | thinking-14lang | official ↗ |
| #36 | Qwen3-0.6B | Qwen | 37.1% | non-thinking-14lang | official ↗ |
Frequently asked questions about MMMLU
What is the MMMLU benchmark?
Multilingual MMLU professionally translated into 14 languages.
How is the MMMLU benchmark scored?
MMMLU 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 MMMLU?
As of the latest reported scores on GenAIList, Claude Opus 4.5 achieves the highest result on MMMLU with a score of 90.8%.
Is a higher MMMLU score better?
Yes. On MMMLU a higher score indicates better performance, so models near the top of the leaderboard are the strongest.