Benchmark
MMMU-Pro
Vision benchmarks measure how well models understand images. Scores are percentages. Higher is better.
- Models and systems measured
- 77
- Scores recorded
- 135
- Sources
- 3
- IQ weight
- 2.51
Score distribution
- 73.5%
- Mean
- 76.9%
- Median
- 11.3%
- Std dev
- 59.9%
- 10th percentile
- 82.3%
- 90th percentile
- 32.7%
- Min
This benchmark has a weight of 2.51 for chat models, measured over 75 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.
Leaderboard
| # | Model or system | Score |
|---|---|---|
| 101 | Gemma 4 26B A4BGemma 4 Technical Report · Table 12 · thinking · 280 vision tokens· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 73.2% |
| 102 | Seed 2.0 Mini· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 71.4% |
| 103 | Claude Opus 4.5Opus 4.6 System Card capability summary · adaptive thinking · max effort · default sampling · 5 trials unless noted · no tools· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 70.6% |
| 104 | GPT 5.4 Nanoxhigh · with Python· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 69.5% |
| 105 | Qwen3 VL 235Bthinking· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. | 69.3% |
| 106 | Gemma 4 12B It· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 69.1% |
| 107 | Claude Sonnet 4.5Opus 4.6 System Card capability summary · adaptive thinking · max effort · default sampling · 5 trials unless noted · with tools· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 68.9% |
| 108 | Qwen3.5 Omni FlashQwen3.5-Omni release · published setting· Reported by Qwen, Qwen Team· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 68.2% |
| 109 | Qwen3 VL 235B· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. | 68.1% |
| 110 | Qwen3 VL 32B Thinking· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. | 68.1% |
| 111 | GPT 5.4 Nano· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 66.1% |
| 112 | GPT 5.4 Nanoxhigh · without Python· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 66.1% |
| 113 | Qwen3 VL 32B Instruct· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. | 65.3% |
| 114 | Claude Sonnet 4.5Opus 4.6 System Card capability summary · adaptive thinking · max effort · default sampling · 5 trials unless noted · no tools· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 63.4% |
| 115 | Command A Plus 05 2026· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 63% |
| 116 | qwen3-vl-30bthinking· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. | 63% |
| 117 | Nova 2 Lite· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 61.8% |
| 118 | Nova 2 LiteAmazon Nova 2 Lite official launch chart / technical report · accuracy, v1· Reported by Amazon Web Services, Amazon AGI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 61.8% |
| 119 | Qwen: Qwen3 VL 8B Thinking· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. | 60.4% |
| 120 | qwen3-vl-30b· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. | 60.4% |
| 121 | Mistral Smalllatest· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 60% |
| 122 | GPT 4o· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 59.9% |
| 123 | Llama 4 Maverick· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 59.6% |
| 124 | Qwen3 VL 8B Instruct· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. | 55.9% |
| 125 | Gemma 4 E4B-it· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 52.6% |
| 126 | Llama 4 ScoutStandard/Vision average· Reported by Meta· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 52.2% |
| 127 | Qwen2.5 VL 72B Instruct· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 51.1% |
| 128 | Qwen2-VL-72B· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 46.2% |
| 129 | Gemma 4 E2B-it· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 44.2% |
| 130 | Phi 4 Multimodal Instruct· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 38.5% |
| 131 | Qwen2.5-VL 7B Instruct· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claimValues from 2 sources
| 38.3% |
| 132 | Qwen2.5-Omni 7B· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 36.6% |
| 133 | Llama 3.2 11B Instruct· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 33% |
| 134 | qwen2.5-vl-3b-instruct· Hugging Face Official Benchmarks data· Reported by the model's developer. The source relaying it did not verify it. claim | 32.7% |
Showing 101–134 of 134 results.
Not ranked
The source marks these results as descriptive, not capability scores, so we list them by name instead of ranking them.
| Model or system | Score |
|---|---|
| DeepSeek V4.1 Flash BaseEM · 4-shot· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 56.5% |
Data credits: BenchBoard by SYuan03 (github.com/SYuan03/benchboard). MIT covers its code and data organization; no results licence is inferred. Numeric relays retain the original stating-source links. · Data by LLM Stats · Hugging Face official benchmark leaderboards; licence varies by benchmark