OpenRef: leaderboard

OpenRef (2026) tests open-world referring expression comprehension: expressions name landmarks, celebrities or logos, use polysemous or ordinal words, and may refer to one or several objects. The score is grounding F1 over predicted and ground-truth boxes in percent, averaged over its categories; higher is better.

Metric: Grounding F1 (%) over predicted and ground-truth boxes, average over the positive-expression categories, 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 8B63.7#917
2Qwen 3 VL 4B60.9#1151
3Qwen 2.5 VL 32B59.4#941
4GLM-4.6V-Flash58.9#1040
5Qwen 3 VL 2B54.2#2134
6Qwen 2.5 VL 7B52.6#1862
7MiniCPM-V-4.548.8#1441
8InternVL3.5-8B40.8#1604
9Keye-VL-1.5-8B40.7#1720
10MiMo-VL-7B-RL37.6#1155
11LLaVA-OneVision-1.5-8B35.3#1864
12Qwen 2 VL 7B34.1#2275
13OpenRef Mistral-3 8B (checkpoint unspecified)5.9

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

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