Molmo2-VideoCount: leaderboard
Video counting benchmark from the Molmo2 paper (Allen Institute for AI and University of Washington, 2026): 533 human-annotated, double-checked validation queries ask how many objects, actions or animals appear in a video clip, with counts up to 60. The score is exact-count accuracy in percent; higher is better.
Metric: Exact-count accuracy (%; 533 validation queries). Source: arxiv.org. Saturation forecast: Around 2031. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 37.1 | #135 |
| 2 | Gemini 2.5 Pro | 35.8 | #262 |
| 3 | GPT-5 | 35.8 | #163 |
| 4 | Molmo2-8B | 35.5 | #1488 |
| 5 | Gemini 2.5 Flash | 31.9 | #452 |
| 6 | GPT-5 Mini | 29.8 | #349 |
| 7 | Claude Sonnet 4.5 | 27.2 | #244 |
| 8 | GLM-4.1V-9B (Thinking) | 26.6 | #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-videocount · How It Works · Data refreshed daily, snapshot 2026-11-12.