OpenRef - Multi Target: leaderboard

The multi target split of OpenRef (2026), an open-world referring expression benchmark: expressions that refer to several objects, to be grounded with bounding boxes in the image. The score is F1 over predicted and ground-truth boxes in percent; higher is better.

Metric: Grounding F1 (%) on expressions that refer to several objects, zero-shot box grounding, each model prompted in its own grounding format; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 13 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 2.5 VL 32B37.4#941
2Qwen 3 VL 8B35.5#917
3Qwen 3 VL 4B33.7#1151
4Qwen 3 VL 2B29.6#2134
5Qwen 2.5 VL 7B28.1#1862
6GLM-4.6V-Flash21.1#1040
7Qwen 2 VL 7B16.8#2275
8MiMo-VL-7B-RL15#1155
9Keye-VL-1.5-8B13.3#1720
10MiniCPM-V-4.511.8#1441
11InternVL3.5-8B9.4#1604
12LLaVA-OneVision-1.5-8B8.7#1864
13OpenRef Mistral-3 8B (checkpoint unspecified)0.2

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

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