OpenRef - Single Target: leaderboard

The single target split of OpenRef (2026), an open-world referring expression benchmark: expressions that refer to one object, 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 one object, 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 3 VL 4B75.2#1151
2Qwen 3 VL 8B74.7#917
3GLM-4.6V-Flash72.9#1040
4Qwen 3 VL 2B66.6#2134
5Qwen 2.5 VL 7B61.8#1862
6Qwen 2.5 VL 32B61.8#941
7MiniCPM-V-4.553.5#1441
8InternVL3.5-8B36#1604
9MiMo-VL-7B-RL27.7#1155
10Keye-VL-1.5-8B26.4#1720
11LLaVA-OneVision-1.5-8B19.1#1864
12Qwen 2 VL 7B16.7#2275
13OpenRef Mistral-3 8B (checkpoint unspecified)0.7

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

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