HarmVideoBench - Clip-Internal Meaning: leaderboard

Metric: Clip-Internal Meaning accuracy (%) on four-option multiple-choice questions about 1,379 real harmful videos (one question per video per level), temperature 0 greedy decoding; API models take the video and audio natively, open models 12 sampled key frames plus the transcript; chance 25; questions about the implied meaning within the clip; higher is better. Source: arxiv.org. Saturation forecast: Around December 2026. 19 models tracked.

Top models

#ModelScore
1Gemini 3.1 Pro (Preview)84.6
2GPT-5.583.4
3Claude Opus 4.782.9
4Qwen 3 VL 235B A22B (Thinking)82.6
5Claude Sonnet 4.682.5
6Qwen 3 VL 235B A22B Instruct82.3
7Qwen 3 VL 32B (Thinking)82.1
8Qwen 3 VL 32B Instruct81.6
9Qwen 3 VL 30B A3B (Thinking)81.4
10Qwen 3 VL 8B (Thinking)79.4
11Qwen 3 VL 30B A3B Instruct78.9
12LFM2.5-VL-1.6B69.3
13GLM-4.5V (Non-reasoning)67.8
14Qwen 3 VL 8B Instruct63.7

Interactive version: theaggregate.ai/benchmark?slug=harmvideobench-clip-internal-meaning · How It Works · Data refreshed daily, snapshot 2026-09-29.