Skip to content
knowledge benchmark · included in ECI

MMLU

MMLU benchmark (score column: EM).
Results
68
Random baseline
25.0%
Score ceiling
100%
Released
7 Sep 2020

Score by model release date

Best score by organisation

#ModelScoreRelativeEvidence
1DeepSeek V3 Open source
DeepSeek
87.2%Source ↗
2Llama 3.3 70B Open source
Meta AI
86.3%Source ↗
3Qwen2.5-72B Open source
Alibaba
85.3%Source ↗
4Phi 4 Open source
Microsoft
84.8%Source ↗
5Llama 3.1-405B Open source
Meta AI
84.5%Source ↗
6Qwen2-72B Open source
Alibaba
82.4%Source ↗
7Llama 3.2 90B Open source
Meta AI
80.3%Source ↗
8Llama 3.1-70B Open source
Meta AI
80.1%Source ↗
9Mistral Large 2 (Jul 2024) Open source
Mistral AI
80.0%Source ↗
10Qwen2.5-14B Open source
Alibaba
79.9%Source ↗
11Llama 3-70B Open source
Meta AI
79.3%Source ↗
12Qwen2.5-Coder-32B Open source
Alibaba
79.1%Source ↗
13DeepSeek-V2 (MoE-236B, May 2024) Open source
DeepSeek
78.4%Source ↗
14phi-3-medium 14B Open source
Microsoft
78.0%Source ↗
15Mixtral 8x22B Open source
Mistral AI
77.8%Source ↗
16Yi-34B Open source
01.AI
76.3%Source ↗
17Gemma 2 27B Open source
Google
75.7%Source ↗
18phi-3-small 7.4B Open source
Microsoft
75.7%Source ↗
19Qwen1.5-32B Open source
Alibaba
74.4%Source ↗
20Qwen2.5-7B Open source
Alibaba
72.9%Source ↗
21Gemma 2 9B Open source
Google DeepMind
72.1%Source ↗
22Mixtral 8x7B Open source
Mistral AI
70.6%Source ↗
23Falcon-180B Open source
Technology Innovation Institute
70.6%Source ↗
24Llama 2-70B Open source
Meta AI
69.9%Source ↗
25Command R+ Open source
Cohere,Cohere Labs (formerly Cohere for AI)
69.4%Source ↗
26phi-3-mini 3.8B Open source
Microsoft
68.8%Source ↗
27Llama 3-8B Open source
Meta AI
68.8%Source ↗
28Qwen1.5-14B Open source
Alibaba
68.6%Source ↗
29Stable Beluga 2 Open source
Stability AI
68.6%Source ↗
30Qwen2.5-Coder (7B) Open source
Alibaba
68.0%Source ↗
31Qwen-14B Open source
Alibaba
66.3%Source ↗
32Gemma 7B Open source
Google DeepMind
66.1%Source ↗
33StarCoder 2 15B Open source
Hugging Face,ServiceNow,NVIDIA,BigCode
64.1%Source ↗
34Yi 6B Open source
01.AI
64.0%Source ↗
35LLaMA-65B Open source
Meta AI
63.4%Source ↗
36Qwen1.5-7B Open source
Alibaba
62.6%Source ↗
37Mistral 7B v0.2 Open source
Mistral AI
62.5%Source ↗
38Mistral 7B v0.1 Open source
Mistral AI
62.5%Source ↗
39Mistral 7B v0.3 Open source
Mistral AI
59.9%Source ↗
40Baichuan2-13B Open source
Baichuan
59.2%Source ↗
41LLaMA-33B Open source
Meta AI
58.7%Source ↗
42Falcon 2 11B Open source
Technology Innovation Institute
58.4%Source ↗
43Phi-2 Open source
Microsoft
58.4%Source ↗
44Falcon-40B Open source
Technology Innovation Institute
56.9%Source ↗
45Llama 3.2 11B Open source
Meta AI
56.5%Source ↗
46Llama 3.1-8B Open source
Meta AI
56.1%Source ↗
47Llama 2-13B Open source
Meta AI
55.6%Source ↗
48Baichuan 2-7B Open source
Baichuan
54.2%Source ↗
49Qwen2.5-Coder (1.5B) Open source
Alibaba
53.6%Source ↗
50Baichuan 1-13B Open source
Baichuan
51.6%Source ↗
51MPT-30B Open source
MosaicML
47.9%Source ↗
52LLaMA-13B Open source
Meta AI
47.7%Source ↗
53Llama 2-7B Open source
Meta AI
45.8%Source ↗
54Qwen-7B Open source
Alibaba
45.0%Source ↗
55Gemma 2B Open source
Google DeepMind
42.3%Source ↗
56Baichuan1-7B Open source
Baichuan
42.3%Source ↗
57DeepSeek Coder 33B Open source
DeepSeek,Peking University
39.4%Source ↗
58StarCoder 2 7B Open source
Hugging Face,ServiceNow,NVIDIA,BigCode
38.8%Source ↗
59Phi-1.5 Open source
Microsoft
37.6%Source ↗
60StarCoder 2 3B Open source
Hugging Face,ServiceNow,NVIDIA,BigCode
36.6%Source ↗
61DeepSeek Coder 6.7B Open source
DeepSeek,Peking University
36.4%Source ↗
62XGen-7B Open source
Salesforce
36.3%Source ↗
63LLaMA-7B Open source
Meta AI
35.6%Source ↗
64Falcon-7B Open source
Technology Innovation Institute
35.0%Source ↗
65MPT-7B Open source
MosaicML
30.8%Source ↗
66Dolly 2.0-12b Open source
Databricks
26.2%Source ↗
67Cerebras-GPT-13B Open source
Cerebras Systems
26.2%Source ↗
68DeepSeek Coder 1.3B Open source
DeepSeek,Peking University
25.8%Source ↗

Caveats: settings (reasoning effort, agent scaffold) differ between rows and materially affect scores; the best reported setting per model is shown. Source: Epoch AI, CC BY 4.0.