VideoEval-Pro: leaderboard

Each model watches uniformly sampled frames of long videos (10 minutes and over, 38 minutes on average) and gives a short free-text answer to each of 1,289 questions; the score is the share a GPT-4o-0806 judge marks correct, at the model's best frame count; the rows are the paper's runs. The questions are Video-MME, MLVU, LVBench and LongVideoBench multiple-choice items rewritten without their options, so guessing among choices is not possible.

Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 16 models tracked. Updated 17 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-4.140.8#505
2GPT-4.1 Mini39.9#793
3Gemini 2.5 Flash (Preview 04-17)Non-reasoning36.3#455
4Qwen 2.5 VL 7B Instruct27.7#1893
5InternVideo2.5-Chat-8B27.2#2031
6VideoChat-Flash-Qwen2-7B (res448)27
7Qwen 2 VL 7B Instruct26.5#2594
8InternVL3-8B24.7#1756
9InternVL2.5-8B24.6#2401
10LLaVA-Video-7B-Qwen224.2#1910
11LongVU-7B22.1
12Video-XL18.6#2583
13Phi-4 Multimodal Instruct16.5#4209
14LongVA-7B16.5#3414
15Mantis-8B-Idefics214.8#6896

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

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