CrossVid: leaderboard

Each model answers questions that require comparing or combining several videos at once (ten tasks; 9,015 questions over 5,331 videos); the score is the source's mean accuracy over the ten tasks. Single-video benchmarks such as Video-MME ask about one clip.

Metric: Overall accuracy (%). Source: raw.githubusercontent.com. Saturation forecast: Around February 2027. 15 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1Gemini 2.5 Pro50.4#274
2GLM-4.1V-9B (Thinking)35.1
3Qwen 2.5 VL 32B Instruct33.7#1074
4Kimi-VL-A3B (Thinking)28.2
5LLaVA-Video-72B-Qwen227.5#1344
6InternVL3-78B25.8#755
7InternVL3-8B25.6#1756
8MiniCPM-o-2.625.6#2417
9Qwen2.5-Omni-7B24.6#1883
10InternVL3-38B23.5#927
11Video-R1-7B21.6#2053
12Phi-3.5-Vision Instruct21.5#6271
13Qwen 2.5 VL 7B Instruct18.3#1893
14LongVA-7B-DPO18#2177
15VideoLLaMA3-7B15.3#2015

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=crossvid · How It Works · Data refreshed daily, snapshot 2026-11-17.