Molmo2-CapTest: leaderboard
Dense video captioning test from the Molmo2 paper (Clark, Zhang, Ma et al., Allen Institute for AI and University of Washington, 2026): models caption 693 videos, and an LLM judge checks the statements in each caption against statements from at least four human captions of the same video. The score is caption F1 in percent, the harmonic mean of statement precision and recall; higher is better.
Metric: Caption F1 (%; LLM-judged statement precision and recall against human captions, 693 videos). Source: arxiv.org. Saturation forecast: Around October 2028. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 Mini | 56.6 | #349 |
| 2 | GPT-5 | 50.1 | #163 |
| 3 | Gemini 2.5 Flash | 46 | #452 |
| 4 | Molmo2-8B | 43.2 | #1488 |
| 5 | Gemini 2.5 Pro | 42.1 | #262 |
| 6 | Gemini 3 Pro | 36 | #135 |
| 7 | Claude Sonnet 4.5 | 26 | #244 |
| 8 | GLM-4.1V-9B (Thinking) | 18.4 | #1061 (GLM-4.1V-9B) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=molmo2-captest · How It Works · Data refreshed daily, snapshot 2026-11-12.