Atlas

Benchmarks

← All benchmarks

MVBench

Multimodal · 2023-11-28

MVBench covers 20 temporal video-understanding tasks, each with 200 multiple-choice questions (4,000 in total) generated from public video datasets, that cannot be solved from a single frame. Scores report average accuracy.

Top models (higher is better)

ModelScore
Qwen3.5-Omni-Plus79.0
GPT-5.278.1
Qwen3.5 397B A17B77.6
Qwen3.5 Plus (2026-02-15)76.7
Qwen3.5 122B A10B76.6
Qwen3-VL-235B-A22B-Instruct76.5
InternVL2.5-78B76.4
Qwen3.6 27B75.5
GPT-575.3
Qwen3 VL 235B A22B Thinking75.2
Qwen3.5 35B A3B74.8
Qwen3.5-27B74.6
Qwen3.6 35B A3B74.6
Qwen3.5-9B74.4
Gemini 3 Pro Preview74.1
Qwen2 VL 72B Instruct73.6
Kimi K2.573.5
Qwen3 VL 32B Thinking73.2
Qwen3 VL 32B Instruct72.8
Qwen3 VL 30B A3B Instruct72.3
Qwen3 VL 30B A3B Thinking72.0
Qwen3.5-4B71.3
Qwen3.5-Omni-Flash70.8
Qwen2.5-VL-72B-Instruct70.4
Gemini 2.5 Pro69.9
Qwen2.5-VL-7B-Instruct69.6
LFM3-VL-2B69.4
Qwen3 VL 4B Thinking69.3
Qwen3 VL 8B Thinking69.0
Qwen3 VL 4B Instruct68.9
Qwen3-VL-8B-Instruct68.7
GPT-5.6 Luna68.2
Opus 4.567.2
Qwen2.5-VL-3B-Instruct67.0
Qwen2 VL 7B Instruct67.0
Gemini 3.5 Flash-Lite64.5
Qwen3 VL 2B Thinking64.5
Qwen2 VL 2B Instruct63.2
Qwen3 VL 2B Instruct61.7
Opus 4.161.4
Loading Atlas data…