Vinoground
Models match short videos with captions that distinguish events by their temporal order.
The benchmark pairs videos with similar objects and different event sequences.
- Made by
- Jianrui Zhang and colleagues
- Released
- October 2024
- Announced on X
- 17K views, 128 likes
- Kind
- Video understanding
- Level
- School
- Scored
- Checked automatically
- Answers
- Prediction
- Tasks
- All public
- Measures
- Capability
- Language
- English
- Interaction
- Single turn
- Refresh
- One fixed set
- Run by
- Anyone
- In views
- Vision
GPT-4o ( OpenAI, 2024a ) (CoT) ( Wei et al., 2022 ) tops the leaderboard with 35%
Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders.
- 35.0 GPT-4o ( OpenAI, 2024a ) (CoT) ( Wei et al., 2022 )
- 24.6 GPT-4o (snapshot unspecified)
- 21.8 LLaVA-OneVision-Qwen2-72B ( Li et al., 2024a )
- 17.4 Qwen2-VL-72B ( Wang et al., 2024 )
- 15.2 Qwen2-VL-7B ( Wang et al., 2024 )
- 14.6 LLaVA-OneVision-Qwen2-7B ( Li et al., 2024a )
- 12.4 Gemini-1.5-Pro ( Gemini Team, 2024 ) (CoT)
- 11.2 MiniCPM-2.6 ( Yao et al., 2024 )
- 10.6 Claude 3.5 Sonnet ( Anthropic, 2024 )
- 10.2 Gemini-1.5-Pro (snapshot unspecified)
- 9.60 InternLM-XC-2.5
- 9.00 InternLM-XC-2.5 ( Zhang et al., 2024b ) (CoT)
- 8.40 VideoLLaMA2-72B ( Cheng et al., 2024 )
- 6.80 LLaVA-NeXT-Video-7B ( Liu et al., 2024a ) (CoT)
- 6.80 MA-LMM-Vicuna-7B ( He et al., 2024 )
- 6.60 Video-LLaVA-7B ( Lin et al., 2024 )
- 6.20 LLaVA-NeXT-Video-7B
- 6.20 Phi-3.5-Vision ( Microsoft, 2024 )
- 5.20 LLaVA-NeXT-Video-34B ( Liu et al., 2024a ) (CoT)
- 3.80 LLaVA-NeXT-Video-34B
- 1.20 LanguageBind ( Zhu et al., 2024 )
- 1.20 VideoCLIP ( Xu et al., 2021 )
- 0.60 ImageBind ( Girdhar et al., 2023 )
- 1 GPT-4o ( OpenAI, 2024a ) (CoT) ( Wei et al., 2022 ) 32 frames 35%
- 2 GPT-4o (snapshot unspecified) 32 frames 24.6%
- 3 LLaVA-OneVision-Qwen2-72B ( Li et al., 2024a ) 32 frames 21.8%
- 4 Qwen2-VL-72B ( Wang et al., 2024 ) 32 frames 17.4%
- 5 Qwen2-VL-7B ( Wang et al., 2024 ) 4fps frames 15.2%
- 6 LLaVA-OneVision-Qwen2-7B ( Li et al., 2024a ) 16 frames 14.6%
- 7 Gemini-1.5-Pro ( Gemini Team, 2024 ) (CoT) 1fps frames 12.4%
- 8 MiniCPM-2.6 ( Yao et al., 2024 ) 16 frames 11.2%
- 9 Claude 3.5 Sonnet ( Anthropic, 2024 ) 4 frames 10.6%
- 10 Gemini-1.5-Pro (snapshot unspecified) 1fps frames 10.2%
- 11 InternLM-XC-2.5 32/1fps frames 9.60%
- 12 InternLM-XC-2.5 ( Zhang et al., 2024b ) (CoT) 32/1fps frames 9.00%
- 13 VideoLLaMA2-72B ( Cheng et al., 2024 ) 8 frames 8.40%
- 14 LLaVA-NeXT-Video-7B ( Liu et al., 2024a ) (CoT) 32 frames 6.80%
- 15 MA-LMM-Vicuna-7B ( He et al., 2024 ) 4 frames 6.80%
- 16 Video-LLaVA-7B ( Lin et al., 2024 ) 8 frames 6.60%
- 17 LLaVA-NeXT-Video-7B 32 frames 6.20%
- 18 Phi-3.5-Vision ( Microsoft, 2024 ) 16 frames 6.20%
- 19 LLaVA-NeXT-Video-34B ( Liu et al., 2024a ) (CoT) 32 frames 5.20%
- 20 LLaVA-NeXT-Video-34B 32 frames 3.80%
- 21 LanguageBind ( Zhu et al., 2024 ) 8 frames 1.20%
- 22 VideoCLIP ( Xu et al., 2021 ) 60 frames 1.20%
- 23 ImageBind ( Girdhar et al., 2023 ) 20 frames 0.60%
Source: The Vinoground paper
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-4o ( OpenAI, 2024a ) (CoT) ( Wei et al., 2022 ) | 35% | 32 frames | Official leaderboard | |
| 2 | GPT-4o (snapshot unspecified) | 24.6% | 32 frames | Official leaderboard | |
| 3 | LLaVA-OneVision-Qwen2-72B ( Li et al., 2024a ) | 21.8% | 32 frames | Official leaderboard | |
| 4 | Qwen2-VL-72B ( Wang et al., 2024 ) | 17.4% | 32 frames | Official leaderboard | |
| 5 | Qwen2-VL-7B ( Wang et al., 2024 ) | 15.2% | 4fps frames | Official leaderboard | |
| 6 | LLaVA-OneVision-Qwen2-7B ( Li et al., 2024a ) | 14.6% | 16 frames | Official leaderboard | |
| 7 | Gemini-1.5-Pro ( Gemini Team, 2024 ) (CoT) | 12.4% | 1fps frames | Official leaderboard | |
| 8 | MiniCPM-2.6 ( Yao et al., 2024 ) | 11.2% | 16 frames | Official leaderboard | |
| 9 | Claude 3.5 Sonnet ( Anthropic, 2024 ) | 10.6% | 4 frames | Official leaderboard | |
| 10 | Gemini-1.5-Pro (snapshot unspecified) | 10.2% | 1fps frames | Official leaderboard | |
| 11 | InternLM-XC-2.5 | 9.60% | 32/1fps frames | Official leaderboard | |
| 12 | InternLM-XC-2.5 ( Zhang et al., 2024b ) (CoT) | 9.00% | 32/1fps frames | Official leaderboard | |
| 13 | VideoLLaMA2-72B ( Cheng et al., 2024 ) | 8.40% | 8 frames | Official leaderboard | |
| 14 | LLaVA-NeXT-Video-7B ( Liu et al., 2024a ) (CoT) | 6.80% | 32 frames | Official leaderboard | |
| 15 | MA-LMM-Vicuna-7B ( He et al., 2024 ) | 6.80% | 4 frames | Official leaderboard | |
| 16 | Video-LLaVA-7B ( Lin et al., 2024 ) | 6.60% | 8 frames | Official leaderboard | |
| 17 | LLaVA-NeXT-Video-7B | 6.20% | 32 frames | Official leaderboard | |
| 18 | Phi-3.5-Vision ( Microsoft, 2024 ) | 6.20% | 16 frames | Official leaderboard | |
| 19 | LLaVA-NeXT-Video-34B ( Liu et al., 2024a ) (CoT) | 5.20% | 32 frames | Official leaderboard | |
| 20 | LLaVA-NeXT-Video-34B | 3.80% | 32 frames | Official leaderboard | |
| 21 | LanguageBind ( Zhu et al., 2024 ) | 1.20% | 8 frames | Official leaderboard | |
| 22 | VideoCLIP ( Xu et al., 2021 ) | 1.20% | 60 frames | Official leaderboard | |
| 23 | ImageBind ( Girdhar et al., 2023 ) | 0.60% | 20 frames | Official leaderboard |
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Audio and video
- Video-MMMU In use 300 expert-level lecture videos with questions testing perception, comprehension and adaptation of knowledge. Accuracy.
Qwen 3.8 Max 88.7%
- LVBench In use Hours-long video understanding with multiple-choice questions on reasoning, grounding and summarization.
Gemini 4 Argon 91.7%
- MMVU In use 3,000 expert-annotated questions on videos across 27 disciplines. Multiple-choice and open-ended accuracy.
Qwen 3.8 Max 82.4%
- MLVU In use Multi-task long video understanding (minutes to hours) with multiple-choice and generation tasks. Accuracy.
Qwen 3.8 Max 90.8%
- FLEURS In use Models transcribe read speech in 102 languages built from FLoRes-101 sentences. Model cards usually report word error rate, so lower scores are better.
Phi-4-multimodal 4.00%
- VoiceBench In use Spoken instructions with speaker, environment and content variation. Voice assistant responses scored by GPT judge or rules.
Qwen3.5-Omni-Plus 93.1%
Data audit
All 8 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 2 pass 3 notes
- Pass Independent runs Scores from the benchmark’s paper (23 entries).
- Pass Models scored 23 entries on its leaderboard, none a model we track.
- Note Newest score No scores yet.
- Note Frontier models No models scored yet.
- Note Headroom No scores yet.
Leaderboard import 3 pass
- Pass Importer We read the board on arxiv.org: 23 entries at the last read, none a model we track.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.