FCMBench-Video (Chinese) - Counting: leaderboard
FCMBench-Video (2026) Chinese-language handheld document videos from financial credit workflows, task counting: exact-match accuracy (%) of the de-duplicated count of logical documents shown in the video. Zero-shot; higher is better.
Metric: Counting: exact-match accuracy (%) of the de-duplicated count of logical documents shown in the video, on FCMBench-Video's zh-CN subset (405 composed videos of 20 Chinese financial-credit document types, 5,960 instances); zero-shot with structured outputs; each model's native video serving path (2 FPS sampling for most open models, 0.5 FPS for Ovis2.5, raw video upload for the API models); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 67.09 | #119 |
| 2 | Doubao-Seed-1.6-vision | 54.67 | #361 |
| 3 | Qwen 3 VL 32B Instruct | 41.11 | #606 |
| 4 | Qwen 3.5 27B | 39.33 | #375 |
| 5 | Qwen 3 VL 8B Instruct | 34.74 | #1029 |
| 6 | InternVL3-8B | 28.22 | #1782 |
| 7 | Qwen3 Omni 30B A3B Instruct | 27.19 | #930 |
| 8 | Ovis2.5-9B | 25.41 | #1225 |
| 9 | Kimi-VL-A3B-Instruct | 18.37 | #2253 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=fcmbench-video-chinese-counting · How It Works · Data refreshed daily, snapshot 2026-11-21.