M3-VQA: leaderboard

M3-VQA (CASIA, 2026) is a knowledge-based visual question answering benchmark of multimodal, multi-entity, multi-hop questions that require recognizing several entities and chaining evidence from Wikipedia, scored by the overlap between predicted and gold answer sets. This board is IoU accuracy in percent with only the image and question as input; higher is better.

Metric: IoU accuracy (%) between the predicted and gold answer sets (exact match against Wikidata answer aliases, one-year margin for dates), averaged over the M3-VQA multi-entity, multi-hop knowledge-based visual questions, with only the image and question as input (no external knowledge); higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 16 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 2.5 VL 72B Instruct32.55#841
2InternVL2.5-78B31.28#1031
3Qwen 2.5 VL 32B Instruct28.66#1062
4InternVL2.5-38B28.55#1373
5GPT-4o27.5#751
6DeepSeek-VL224.32#2927
7MiniCPM-V-2.623.45#3267
8Qwen 2.5 VL 7B Instruct22.53#1902
9LLaVA-OneVision-7B22.02#2981
10Qwen 2 VL 7B Instruct21.71#2573
11InternVL2.5-4B18.82#2549
12InternVL2.5-8B18.69#2370
13Qwen 2.5 VL 3B Instruct18.66#2713
14InternVL2.5-26B16.48#2156
15InternVL2.5-2B8.41#6468

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

Interactive version: theaggregate.ai/benchmark?slug=m3-vqa · How It Works · Data refreshed daily, snapshot 2026-11-21.