Benchmark Almanac

RefCOCO / RefCOCO+ / RefCOCOg

Qwen3.5-Omni-Plus 95% leads since Mar 2026 In use

Model grounds a referring expression to a bounding box in a COCO image. Scored by accuracy at IoU 0.5.

Reka's 'RefCOCO-A/B' are the testA/testB splits.

In the authors’ words

Humans refer to objects in their environments all the time, especially in dialogue with other people. We explore generating and comprehending natural language referring expressions for objects in images. In particular, we focus on incorporating better measures of visual context into referring expression models and find that visual comparison to other objects within an image helps improve performance significantly. We also develop methods to tie the language generation process together, so that we generate expressions for all objects of a particular category jointly. Evaluation on three recent datasets - RefCOCO, RefCOCO+, and RefCOCOg, shows the advantages of our methods for both referring expression generation and comprehension.

Abstract of Modeling Context in Referring Expressions (arXiv:1608.00272, CC0)
Made by UNC Chapel Hill (Licheng Yu et al.) Released Jul 2016 Detection and captioning All tasks public Last in a launch post Aug 2026 Last model added Aug 2026
Level
Everyday Most adults, without training
Scored
Checked automatically A program checks the result against an answer key, unit tests or the expected final state.
Answers
Short answer A number, word or short string matched to a key.
lichengunc/refer 577 stars Paper
19
Models scored in lab cards
17
Launch posts list it, from 10 labs
14
System cards report it, from 8 labs

Qwen3.5-Omni-Plus leads with 95% among models released by Aug 2026

Best score per model as labs report them in their own cards. Labs pick their own settings, so these compare less cleanly than independent runs.

The newest model with a score here came out in Aug 2026, and nobody has reported a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

19 models. Scroll sideways for the rest.

  1. 95.0 Qwen3.5-Omni-Plus
  2. 93.9 LongCat-Flash-Omni
  3. 92.5 Qwen3.6 27B
  4. 92.4 InternVL3.5-241B-A28B
  5. 92.4 Qwen3-VL-235B-A22B
  6. 92.0 Qwen 3.6 35B-A3B
  7. 91.9 Intern-S1-Pro
  8. 91.3 Seed1.5-VL
  9. 91.3 GLM-4.5V
  10. 91.3 Qwen3.5-122B-A10B
  11. 90.5 Nemotron 3 Nano Omni 30B-A3B
  12. 89.9 Reka Edge
  13. 88.6 GLM-4.6V
  14. 88.1 Keye-VL-2.0-30B-A3B
  15. 87.9 LFM2.5-VL-3B
  16. 87.7 Qwen2.5-Omni-7B
  17. 87.4 GLM-4.1V-9B-Thinking
  18. 86.3 Nova 2 Omni Preview
  19. 73.2 North Micro Vision Instruct
  1. 1 Qwen3.5-Omni-Plus 95%
  2. 2 LongCat-Flash-Omni 93.9%
  3. 3 Qwen3.6 27B 92.5%
  4. 4 InternVL3.5-241B-A28B 92.4%
  5. 5 Qwen3-VL-235B-A22B 92.4%
  6. 6 Qwen 3.6 35B-A3B 92%
  7. 7 Intern-S1-Pro 91.9%
  8. 8 Seed1.5-VL 91.3%
  9. 9 GLM-4.5V 91.3%
  10. 10 Qwen3.5-122B-A10B 91.3%
  11. 11 Nemotron 3 Nano Omni 30B-A3B 90.5%
  12. 12 Reka Edge 89.9%
  13. 13 GLM-4.6V 88.6%
  14. 14 Keye-VL-2.0-30B-A3B 88.1%
  15. 15 LFM2.5-VL-3B 87.9%
  16. 16 Qwen2.5-Omni-7B 87.7%
  17. 17 GLM-4.1V-9B-Thinking 87.4%
  18. 18 Nova 2 Omni Preview 86.3%
  19. 19 North Micro Vision Instruct 73.2%
#ModelScoreSettingDateSource
1 Qwen3.5-Omni-Plus 95% Mar 2026 Model card
2 LongCat-Flash-Omni 93.9% Oct 2025 Model card
3 Qwen3.6 27B 92.5% Apr 2026 Model card
4 InternVL3.5-241B-A28B 92.4% Aug 2025 Model card
5 Qwen3-VL-235B-A22B 92.4% Sep 2025 Model card
6 Qwen 3.6 35B-A3B 92% Apr 2026 Model card
7 Intern-S1-Pro 91.9% Feb 2026 Model card
8 Seed1.5-VL 91.3% May 2025 Model card
9 GLM-4.5V 91.3% Aug 2025 Model card
10 Qwen3.5-122B-A10B 91.3% Feb 2026 Model card
11 Nemotron 3 Nano Omni 30B-A3B 90.5% Apr 2026 Model card
12 Reka Edge 89.9% Mar 2026 Model card
13 GLM-4.6V 88.6% Dec 2025 Model card
14 Keye-VL-2.0-30B-A3B 88.1% May 2026 Model card
15 LFM2.5-VL-3B 87.9% Aug 2026 Model card
16 Qwen2.5-Omni-7B 87.7% Mar 2025 Model card
17 GLM-4.1V-9B-Thinking 87.4% Jul 2025 Model card
18 Nova 2 Omni Preview 86.3% Dec 2025 Model card
19 North Micro Vision Instruct 73.2% Aug 2026 Model card

5 models have held number one since 2023

Lab claims count from the day the card came out, independent runs from the model’s release.

Qwen ByteDance Seed Shanghai AI Lab Meituan
Lab claims Passed 85% in Mar 2025, by Qwen2.5-Omni-7B
Qwen3.5-Omni-Plus 95%

The best score passed 85% in Mar 2025

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Reported by the lab in its model card Best so far
0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) Qwen3.5-Omni-Plus LongCat-Flash-Omni Seed1.5-VL 0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) Qwen3.5-Omni-Plus Qwen2.5-Omni-7B

2 labs put it in a launch post in the last six months

Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.

Qwen
ByteDance Seed
Z.ai
Meituan
Amazon
Shanghai AI Lab
Reka
Kuaishou
Cohere
Liquid AI
DateLaunch postScoreAs the post writes it
12 Aug 2026 LFM2.5-VL-3B 87.9 RefCOCO-avg
12 Aug 2026 North Micro Vision Instruct 73.2 RefCOCO avg (average of 8 RefCOCO, RefCOCO+ and RefCOCOg splits)
25 May 2026 Keye-VL-2.0-30B-A3B 88.1 RefCOCO (avg)
22 Apr 2026 Qwen3.6 27B 92.5 RefCOCO avg
15 Apr 2026 Qwen 3.6 35B-A3B 92 RefCOCO(avg)
30 Mar 2026 Qwen3.5-Omni-Plus 95 RefCOCO(avg)
11 Mar 2026 Reka Edge 89.9 Mean of RefCOCO-A 93.13 / RefCOCO-B 86.70
24 Feb 2026 Qwen3.5-122B-A10B 91.3 RefCOCO(avg)
2 Feb 2026 Intern-S1-Pro 91.9 RefCOCO-avg
8 Dec 2025 GLM-4.6V 88.6 RefCOCO-avg (val)
2 Dec 2025 Nova 2 Omni Preview 86.3 RefCOCOg acc@0.5
31 Oct 2025 LongCat-Flash-Omni 93.9 RefCOCO-avg
23 Sep 2025 Qwen3-VL-235B-A22B 92.4 RefCOCO-avg
11 Aug 2025 GLM-4.5V 91.3 RefCOCO-avg (val)
1 Jul 2025 GLM-4.1V-9B-Thinking 87.4 RefCOCO-avg (val)
12 May 2025 Seed1.5-VL 91.3 RefCOCO-avg
27 Mar 2025 Qwen2.5-Omni-7B 87.7 Mean of Refcoco val/testA/testB, Refcoco+ val/testA/testB, Refcocog val/test (90.5/93.5/86.6/85.4/91.0/79.3/87.4/87.9)

14 system cards report it, from 8 labs

Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.

DateDocumentScoreAs the document writes itSection
9 Jun 2026 Keye-VL-2.0-30B-A3B technical report 88.1 RefCOCO (avg) (Figure 5, printed value) Capabilities
27 Apr 2026 Nemotron 3 Nano Omni 30B-A3B technical report 90.5 RefCOCO, reasoning on, 80.6 reasoning off (Table 7) Capabilities
17 Apr 2026 Qwen3.5-Omni-Plus technical report 95 RefCOCO(avg) (Table 6) Capabilities
26 Mar 2026 Intern-S1-Pro technical report 91.9 RefCOCO-avg, non-thinking (Table 2) Capabilities
1 Jan 2026 GLM-4.6V technical report 88.6 RefCOCO-avg (val), GLM-4.6V thinking (Table 2) Capabilities
2 Dec 2025 Nova 2 Omni Preview technical report 86.3 RefCOCOg acc@0.5 (Table 6) Capabilities
26 Nov 2025 Qwen3-VL-235B-A22B technical report 92.1 RefCOCO-avg (Table 2) Capabilities
31 Oct 2025 LongCat-Flash-Omni technical report 93.9 RefCOCO-avg, average of RefCOCO and RefCOCO+ (Table 5) Capabilities
25 Aug 2025 InternVL3.5-241B-A28B technical report 92.4 RefCOCO, RefCOCO+ and RefCOCOg Overall (Table 7). RefCOCO val 94.1, testA 96.3, testB 91.5. RefCOCO+ val 91.6, testA 94.6, testB 86.9. RefCOCOg val 92.0, test 92.1 Capabilities
13 Aug 2025 GLM-4.5V technical report 91.3 RefCOCO-avg (val), GLM-4.5V thinking (Table 2) Capabilities
1 Jul 2025 GLM-4.1V-9B-Thinking technical report 85.3 RefCOCO-avg (val), GLM-4.1V-9B-Thinking thinking (Table 2). Table 3 prints 87.4 Capabilities
11 May 2025 Seed1.5-VL technical report 91.3 RefCOCO-avg, thinking (Table 6). Non-thinking 91.6 Capabilities
26 Mar 2025 Qwen2.5-Omni-7B technical report Refcoco val 90.5, testA 93.5, testB 86.6. Refcoco+ val 85.4, testA 91.0, testB 79.3. Refcocog val 87.4, test 87.9. No overall score (Table 6) Capabilities
19 Feb 2025 Qwen2.5-VL-72B technical report RefCOCO val 92.7, testA 94.6, testB 89.7, RefCOCO+ val 88.9, testA 92.2, testB 83.7, RefCOCOg val 89.9, test 90.3, no overall (Table 6) Capabilities

More in Vision

All 234 in Vision
  • MMMU In use 11.5K college-level multimodal questions across 30 subjects. Multiple-choice and open accuracy. GPT-5.1 85.4%
  • MMMU-Pro In use Harder MMMU: college-level multimodal questions with more options and vision-only inputs to remove text shortcuts. Kimi K3.1 Pro 84.1%
  • CharXiv In use Chart understanding on real arXiv figures. Reasoning questions require synthesizing information from complex charts. Claude Fable 5 93.5%
  • MATH-Vision (MathVision) In use 3,040 competition math problems with visual context. Answer accuracy. Qwen 3.8 Max 97.7%
  • ChartQA In use Models answer human-written and generated questions about chart images that need visual and arithmetic reasoning. Scored by relaxed accuracy with a 5 percent numeric tolerance. Command A Vision 90.9%
  • DocVQA In use Models answer 50,000 questions about 12,000 scanned document images by reading their text and layout. Scoring is ANLS, a string similarity to reference answers. Seed1.5-VL 96.9%