Benchmark

MMMU

Massive Multi-discipline Multimodal Understanding

College-level multimodal reasoning across art, business, science, health, humanities, tech.

Models Tested

Top Model

80.4%

Top Score

100 %

Max Possible

#	Model	MMMU Score	Provider	Composite
1	o3 OpenAI	80.4%	OpenAI	97.1
2	Gemini 2.5 Pro Google	79.5%	Google	97.5
3	Claude Opus 4 Anthropic	78.2%	Anthropic	97.8
4	o4-mini OpenAI	74.8%	OpenAI	91.5
5	Llama 4 Maverick Meta	73.1%	Meta	87.5
6	Grok 3 xAI	72.8%	xAI	88.8
7	Claude Sonnet 4 Anthropic	72.5%	Anthropic	90.3
8	DeepSeek R1 DeepSeek	71.5%	DeepSeek	92.3
9	Gemini 2.5 Flash Google	70.1%	Google	86.7
10	GPT-4o OpenAI	69.1%	OpenAI	89.2
11	Claude 3.5 Sonnet Anthropic	68.3%	Anthropic	88.4
12	Qwen 3 235B Alibaba	66.5%	Alibaba	84.5
13	Llama 3.1 405B Meta	64.5%	Meta	82.1
14	Gemini 2.0 Flash Google	64.3%	Google	80.5
15	DeepSeek V3 DeepSeek	64.2%	DeepSeek	86.1
16	Llama 3.3 70B Meta	62.8%	Meta	79.8
17	Qwen 2.5 72B Alibaba	61.2%	Alibaba	76
18	Claude 3.5 Haiku Anthropic	60.2%	Anthropic	79.2
19	Mistral Large Mistral AI	59.8%	Mistral AI	76.5
20	GPT-4o Mini OpenAI	59.4%	OpenAI	75.8
21	Grok 2 xAI	58.5%	xAI	74.8
22	Llama 3.1 70B Meta	58.2%	Meta	74.3
23	Phi-4 Microsoft	52.3%	Microsoft	64.5
24	Mixtral 8x22B Mistral AI	48.5%	Mistral AI	65.2
25	Command R+ Cohere	45.1%	Cohere	60.8

Source: Yue et al., 2024