Video-R1-7B: benchmark results
Video-R1 project’s final 7B video-reasoning checkpoint, released with weights and datasets on March 28, 2025. It follows the distinct February preliminary version and uses temporal GRPO on a Qwen2.5-VL backbone. Provider: Other. Released 2025-03-28. Access: Open.
Aggregate Rating 1525 ± 42, rank #2014 of 10916 rated models, from 30 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| SceneBench - Clip QA | 71.9 | Accuracy (%) on SceneBench's 368 clip qa items (questions ab | 88.9 |
| RepProbe Real - Normalized MAE | 1.35 | Normalized MAE (lower is better) | 66.7 |
| RepProbe Synthetic - OBO Few | 50.9 | OBO Few (%) | 66.7 |
| SceneBench | 54.5 | Average accuracy (%) over SceneBench's six tasks (title and | 66.7 |
| SceneBench - Inverse VQA | 40.7 | Accuracy (%) on SceneBench's 485 inverse vqa items (choosing | 61.1 |
| SceneBench - Title Prediction | 98.5 | Accuracy (%) on SceneBench's 345 title prediction items (cho | 61.1 |
| RepProbe Real - Exact | 7.2 | Exact (%) | 50 |
| RepProbe Real - MAE | 6.24 | MAE (lower is better) | 50 |
| RepProbe Real - OBO | 19.3 | OBO (%) | 50 |
| RepProbe Real - OBO Mid | 21.4 | OBO Mid (%) | 50 |
| RepProbe Synthetic - OBO | 15.3 | OBO (%) | 50 |
| RepProbe Synthetic - OBO Many | 0 | OBO Many (%) | 50 |
Interactive version: theaggregate.ai/model?slug=video-r1-7b · How It Works · Data refreshed daily, snapshot 2026-11-08.