MMMU Pro Benchmark

MMMU Pro Benchmark | Phaseo

Models Using This Benchmark

Organisation	Model	Reported	Top Score	Info	Self Reported	Source
OpenAI	GPT 5.5	23 Apr 2026	83.20%	With tools	Yes	Source
OpenAI	GPT 5.4	05 Mar 2026	82.10%	With tools	Yes	Source
Moonshot	Kimi K3	16 Jul 2026	81.60%	Average of three runs; official protocol preserves input order and prepends images to text input.	Yes	Source
Google	Gemini 3 Flash Preview	17 Dec 2025	81.20%	-	Yes	Source
OpenAI	GPT 5 Pro	07 Aug 2025	81.20%	inferred family alias from gpt-5.4 (score=0.4083; benches=19)	Yes	Source
OpenAI	GPT 5 Search API	14 Oct 2025	81.20%	inferred family alias from gpt-5.4 (score=0.3050; benches=19)	Yes	Source
Google	Gemini 3 Pro Image Preview (Nano Banana Pro)	20 Nov 2025	81%	inferred modality/version alias from gemini-3-pro-preview	Yes	Source
Google	Gemini 3 Pro Preview	18 Nov 2025	81%	-	Yes	Source
Google	Gemini 3.1 Pro Preview	19 Feb 2026	80.50%	Search + Python + Browse	Yes	-
Meta	Muse Spark	08 Apr 2026	80.40%	Multimodal understanding	Yes	Source
OpenAI	GPT 5.2 Chat	11 Dec 2025	79.50%	inferred alias from gpt-5.2-2025-12-11	Yes	Source
Qwen	Qwen 3.5 397B A17B	16 Feb 2026	79%	-	Yes	Source
Qwen	Qwen 3.6 Plus	01 Apr 2026	78.80%	-	Yes	Source
Moonshot	Kimi K2.5	27 Jan 2026	78.50%	-	Yes	Source
OpenAI	GPT 5	07 Aug 2025	78.40%	With Thinking, Pass @ 1	Yes	Source
ByteDance	Seed 2.0 Pro	14 Feb 2026	78.20%	Seed2 official benchmark table \| MMMU-Pro	Yes	Source
OpenAI	GPT 5.4 Mini	17 Mar 2026	78%	With Python	Yes	Source
Xiaomi	MiMo V2.5	22 Apr 2026	77.90%	Image understanding benchmark from official MiMo-V2.5 model card image	Yes	Source
Anthropic	Claude Opus 4.6	05 Feb 2026	77.30%	With tools	Yes	Source
Google	Gemma 4 31B	02 Apr 2026	76.90%	-	Yes	Source
Qwen	Qwen 3.5 122B A10B	24 Feb 2026	76.90%	-	Yes	Source
Google	Gemini 3.1 Flash Image Preview (Nano Banana 2)	26 Feb 2026	76.80%	inferred modality/version alias from gemini-3.1-flash-lite-preview	Yes	Source
Google	Gemini 3.1 Flash-Lite	07 May 2026	76.80%	-	Yes	Source
Google	Gemini 3.1 Flash Lite Preview	03 Mar 2026	76.80%	-	Yes	Source
Google	Veo 3.1 Lite Preview	31 Mar 2026	76.80%	manual fallback alias from gemini-3.1-flash-lite-preview	Yes	Source
ByteDance	Seed 2.0 Lite	14 Feb 2026	76%	Seed2 official benchmark table \| MMMU-Pro	Yes	Source
Anthropic	Claude Sonnet 4.6	17 Feb 2026	75.60%	With tools	Yes	Source
Qwen	Qwen 3.5 35B A3B	24 Feb 2026	75.10%	-	Yes	Source
Qwen	Qwen 3.5 27B	24 Feb 2026	75%	-	Yes	Source
Qwen	Qwen 3.5 Flash	23 Feb 2026	75%	inferred family alias from qwen3.5-27b (score=0.4147; benches=81)	Yes	Source
OpenAI	GPT 5 Mini	07 Aug 2025	74.10%	High Reasoning Effort	Yes	Source
Google	Gemma 4 26B A4B	02 Apr 2026	73.80%	-	Yes	Source
ByteDance	Seed 2.0 Mini	14 Feb 2026	71.40%	Seed2 official benchmark table \| MMMU-Pro	Yes	Source
OpenAI	GPT 5.4 Nano	17 Mar 2026	69.50%	With Python	Yes	Source
Qwen	Qwen 3 VL 235B A22B Thinking	-	69.30%	-	Yes	-
Google	Gemma 4 12B	03 Jun 2026	69.10%	-	Yes	Source
Qwen	Qwen 3 VL 235B A22B Instruct	-	68.10%	-	Yes	-
Qwen	Qwen 3 VL 32B Thinking	-	68.10%	-	Yes	-
Qwen	Qwen 3 VL 32B Instruct	-	65.30%	-	Yes	-
Amazon	Nova 2 Pro	02 Dec 2025	63.50%	-	Yes	Source
Cohere	Command A+	20 May 2026	63%	-	Yes	Source
Qwen	Qwen 3 VL 30B A3B Thinking	-	63%	-	Yes	-
OpenAI	GPT 5 Nano	07 Aug 2025	62.60%	High Reasoning Effort	Yes	Source
Amazon	Nova 2 Lite	02 Dec 2025	61.80%	-	Yes	Source
Qwen	Qwen 3 8B	-	60.40%	inferred high-confidence family alias from qwen3-vl-8b-thinking (score=0.4600; benches=50)	Yes	-
Qwen	Qwen 3 Embedding 8B	-	60.40%	inferred family alias from qwen3-vl-8b-thinking (score=0.3850; benches=50)	Yes	-
Qwen	Qwen 3 Guard Gen 8B	-	60.40%	inferred family alias from qwen3-vl-8b-thinking (score=0.3400; benches=50)	Yes	-
Qwen	Qwen 3 Guard Stream 8B	-	60.40%	inferred family alias from qwen3-vl-8b-thinking (score=0.3371; benches=50)	Yes	-
Qwen	Qwen 3 Reranker 8B	-	60.40%	inferred family alias from qwen3-vl-8b-thinking (score=0.3850; benches=50)	Yes	-
Qwen	Qwen 3 VL 30B A3B Instruct	-	60.40%	-	Yes	-
Qwen	Qwen 3 VL 8B Thinking	-	60.40%	-	Yes	-
Qwen	Qwen 3 VL Embedding 8B	-	60.40%	inferred high-confidence family alias from qwen3-vl-8b-thinking (score=0.5232; benches=50)	Yes	-
Qwen	Qwen 3 VL Reranker 8B	-	60.40%	inferred high-confidence family alias from qwen3-vl-8b-thinking (score=0.5275; benches=50)	Yes	-
Mistral	Mistral Small 1.0	26 Feb 2024	60%	inferred family alias from mistral-small-latest (score=0.3650; benches=9)	Yes	Source
Mistral	Mistral Small 2.0	17 Sept 2024	60%	inferred family alias from mistral-small-latest (score=0.3650; benches=9)	Yes	Source
Mistral	Mistral Small 4	16 Mar 2026	60%	-	Yes	Source
Mistral	Mistral Small Creative	16 Dec 2025	60%	inferred family alias from mistral-small-latest (score=0.4273; benches=9)	Yes	Source
OpenAI	GPT 4o (2024-08-06)	06 Aug 2024	59.90%	-	Yes	Source
OpenAI	GPT 4o Audio (2025-06-03)	03 Jun 2025	59.90%	inferred modality/version alias from gpt-4o-2024-08-06	Yes	Source
OpenAI	GPT 4o Audio (2024-10-01)	01 Oct 2024	59.90%	inferred modality/version alias from gpt-4o-2024-08-06	Yes	Source
OpenAI	GPT 4o Audio (2024-12-17)	17 Dec 2024	59.90%	inferred modality/version alias from gpt-4o-2024-08-06	Yes	Source
OpenAI	GPT 4o Realtime Preview (2025-06-03)	03 Jun 2025	59.90%	inferred modality/version alias from gpt-4o-2024-08-06	Yes	Source
OpenAI	GPT 4o Realtime Preview (2024-10-01)	01 Oct 2024	59.90%	inferred modality/version alias from gpt-4o-2024-08-06	Yes	Source
OpenAI	GPT 4o Search Preview	11 Mar 2025	59.90%	inferred modality/version alias from gpt-4o-2024-08-06	Yes	Source
OpenAI	GPT 4o Transcribe	20 Mar 2025	59.90%	inferred modality/version alias from gpt-4o-2024-08-06	Yes	Source
OpenAI	GPT 4o Transcribe Diarize	15 Oct 2025	59.90%	inferred modality/version alias from gpt-4o-2024-08-06	Yes	Source
Meta	Llama 4 Maverick	05 Apr 2025	59.60%	-	Yes	Source
Qwen	Qwen 3 4B Instruct 2507	-	57%	inferred family alias from qwen3-vl-4b-thinking (score=0.3462; benches=48)	Yes	-
Qwen	Qwen 3 4B SafeRL	-	57%	inferred family alias from qwen3-vl-4b-thinking (score=0.3850; benches=48)	Yes	-
Qwen	Qwen 3 4B Thinking 2507	-	57%	inferred family alias from qwen3-vl-4b-thinking (score=0.3462; benches=48)	Yes	-
Qwen	Qwen 3 Embedding 4B	-	57%	inferred family alias from qwen3-vl-4b-thinking (score=0.3850; benches=48)	Yes	-
Qwen	Qwen 3 Guard Gen 4B	-	57%	inferred family alias from qwen3-vl-4b-thinking (score=0.3400; benches=48)	Yes	-
Qwen	Qwen 3 Guard Stream 4B	-	57%	inferred family alias from qwen3-vl-4b-thinking (score=0.3371; benches=48)	Yes	-
Qwen	Qwen 3 Reranker 4B	-	57%	inferred family alias from qwen3-vl-4b-thinking (score=0.3850; benches=48)	Yes	-
Qwen	Qwen 3 VL 4B Thinking	-	57%	-	Yes	-
Qwen	Qwen 3 VL 8B Instruct	-	55.90%	-	Yes	-
Google	DiffusionGemma 26B A4B IT	10 Jun 2026	54.30%	-	Yes	Source
Qwen	Qwen 3 VL 4B Instruct	-	53.20%	-	Yes	-
Qwen	Qwen 2.5 Math 72B	-	51.10%	inferred high-confidence family alias from qwen2.5-vl-72b (score=0.4700; benches=30)	Yes	Source
Qwen	Qwen 2.5 Math PRM 72B	-	51.10%	inferred family alias from qwen2.5-vl-72b (score=0.4050; benches=30)	Yes	Source
Qwen	Qwen 2.5 Math RM 72B	-	51.10%	inferred family alias from qwen2.5-vl-72b (score=0.4092; benches=30)	Yes	Source
Qwen	Qwen 2.5 VL 72B Instruct	26 Jan 2025	51.10%	-	Yes	Source
Qwen	Qwen 2.5 Coder 32B Instruct	-	49.50%	inferred high-confidence family alias from qwen2.5-vl-32b (score=0.4641; benches=28)	Yes	Source
Qwen	Qwen 2.5 VL 32B Instruct	24 Mar 2025	49.50%	-	Yes	Source
Qwen	Qwen 2.5 VL 3B Instruct	-	49.50%	inferred high-confidence family alias from qwen2.5-vl-32b (score=0.4914; benches=28)	Yes	Source
Qwen	Qwen 2 VL 2B	-	46.20%	inferred family alias from qwen2-vl-72b (score=0.4083; benches=15)	Yes	Source
Qwen	Qwen 2 VL 72B	-	46.20%	-	Yes	Source
Meta	Llama 3.2 90B Vision Instruct	-	45.20%	inferred high-confidence family alias from llama-3.2-90b-instruct (score=0.5873; benches=13)	Yes	Source
Microsoft	Phi 4 multimodal instruct	01 Feb 2025	38.50%	-	Yes	Source
Qwen	Qwen 2.5 0.5B	-	38.30%	inferred family alias from qwen2.5-vl-7b (score=0.3131; benches=32)	Yes	Source
Qwen	Qwen 2.5 1.5B	-	38.30%	inferred family alias from qwen2.5-vl-7b (score=0.3131; benches=32)	Yes	Source
Qwen	Qwen 2.5 3B	-	38.30%	inferred family alias from qwen2.5-vl-7b (score=0.3531; benches=32)	Yes	Source
Qwen	Qwen 2.5 VL 7B Instruct	-	38.30%	-	Yes	Source
Qwen	Qwen 2.5 Coder 3B	-	36.60%	inferred family alias from qwen2.5-omni-7b (score=0.3000; benches=45)	Yes	Source
Qwen	Qwen 2.5 Coder 7B	-	36.60%	inferred high-confidence family alias from qwen2.5-omni-7b (score=0.4700; benches=45)	Yes	Source
Qwen	Qwen 2.5 Math 7B	-	36.60%	inferred high-confidence family alias from qwen2.5-omni-7b (score=0.4767; benches=45)	Yes	Source
Qwen	Qwen 2.5 Math 7B PRM800K	-	36.60%	inferred family alias from qwen2.5-omni-7b (score=0.3696; benches=45)	Yes	Source
Qwen	Qwen 2.5 Math PRM 7B	-	36.60%	inferred family alias from qwen2.5-omni-7b (score=0.4092; benches=45)	Yes	Source
Qwen	Qwen 2.5 Omni 3B	-	36.60%	inferred high-confidence family alias from qwen2.5-omni-7b (score=0.4933; benches=45)	Yes	Source
Qwen	Qwen 2.5 Omni 7B	-	36.60%	-	Yes	Source
Meta	Llama 3.2 11B Vision Instruct	-	33%	inferred high-confidence family alias from llama-3.2-11b-instruct (score=0.5873; benches=11)	Yes	Source
Meta	Llama 3.2 1B Instruct	25 Sept 2024	33%	inferred family alias from llama-3.2-11b-instruct (score=0.4200; benches=11)	Yes	Source

Average Score

Score Range

Leading Model

MMMU Pro

Recorded Results

Average Score

Score Range

Leading Model

Models Using This Benchmark