Benchmark Almanac

Unicorn (Vision Safety)

Qwen2-VL-72B-Instruct 91.2% leads since August 2024
Not in model cards

Models answer questions about objects in unusual images and sketches. The benchmark tests how well visual judgments hold up when the images change.

Made by
Haoqin Tu and colleagues
Released
November 2023
Kind
General visual questions
Level
Everyday
Scored
Checked automatically
Answers
Short answer
Tasks
All public
Measures
Capability
Language
English
In and out
Text, Images → Text
Interaction
Single turn
Refresh
One fixed set
Run by
Anyone
In views
Vision
Strongest model tested
o3 #84 on the capability index
Last model added
April 2025
Scores refreshed
November 20th, 2026
43
Entries on VHELM’s leaderboard, 41 models we track

Qwen2-VL-72B-Instruct leads with 91.2% among models released by April 2025 Qwen2-VL-72B-Instruct leads with 91.2% among models released by April 2025

Best score per entry from VHELM’s leaderboard, read on November 20th, 2026. VHELM reports prefix quasi-exact match across OODCV-VQA and Sketchy-VQA. Its leaderboard omits their challenge variants. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. This view shows only models we track, each at its best.

The newest model with a score here came out in April 2025, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 1 Qwen2-VL-72B-Instruct 91.2%
  2. 2 Qwen2-VL-7B-Instruct 90.2%
  3. 3 Gemini 2.0 Flash Feb 2025 89.9%
  4. 4 Gemini 1.5 Flash Sep 2024 89.4%
  5. 5 Gemini 1.5 Pro Sep 2024 89%
  6. 6 Gemini 1.5 Flash May 2024, BLOCK_NONE safety 88.8%
  7. 7 Gemini 2.0 Pro 88.6%
  8. 8 GPT-4.5 88.3%
  9. 9 Gemini 2.5 Pro Mar 2025 88%
  10. 10 Gemini 2.0 Flash-Lite 87.9%
  11. 11 Gemini 1.5 Pro May 2024, BLOCK_NONE safety 87.8%
  12. 12 o4-mini High reasoning effort 87.5%
  13. 13 Qwen2.5-VL-7B 87.2%
  14. 14 palmyra-vision-003 87.2%
  15. 15 Qwen2.5-VL-72B 87%
  16. 16 o3 High reasoning effort 87%
  17. 17 GPT-4o Nov 2024 86.8%
  18. 18 Gemini 1.0 Pro Vision 86.3%
  19. 19 GPT-4.1 85.7%
  20. 20 GPT-4.1 mini 85.6%
  21. 21 paligemma-3b-mix-224 85.3%
  22. 22 o1 84.8%
  23. 23 Llama 4 Scout 84.5%
  24. 24 Llama 4 Maverick 84.2%
  25. 25 GPT-4o Aug 2024 83%
  26. 26 PaliGemma-3B-mix-448 82.3%
  27. 27 GPT-4o mini 82.2%
  28. 28 Qwen2.5-VL-32B 81.8%
  29. 29 GPT-4o May 2024 81.5%
  30. 30 GPT-4.1 nano 81.3%
  31. 31 GPT-4 Turbo Apr 2024 81.2%
  32. 32 Qwen2.5-VL-3B 80.8%
  33. 33 GPT-4 Turbo Nov 2023 79.6%
  34. 34 Pixtral 12B 63.5%
  35. 35 Claude 3.5 Sonnet (October 2024) 63%
  36. 36 Idefics2-8B 62.9%
  37. 37 Claude 3.5 Sonnet 62.1%
  38. 38 IDEFICS-9B-Instruct 60%
  39. 39 Claude 3 Sonnet 59.5%
  40. 40 Claude 3 Haiku 58.5%
  41. 41 IDEFICS-80B-Instruct 54%
  42. 42 Claude 3 Opus 52.5%
  43. 43 Qwen-VL-Chat 48.8%

Source: VHELM’s leaderboard

Source: VHELM’s leaderboard

#EntryScoreSettingDateSource
1 Qwen2-VL-72B-Instruct 91.2% Aug 2024 Official leaderboard
2 Qwen2-VL-7B-Instruct 90.2% Aug 2024 Official leaderboard
3 Gemini 2.0 Flash Feb 2025 89.9% Feb 2025 Official leaderboard
4 Gemini 1.5 Flash Sep 2024 89.4% Aug 2024 Official leaderboard
5 Gemini 1.5 Pro Sep 2024 89% Sep 2024 Official leaderboard
6 Gemini 1.5 Flash May 2024 88.8% BLOCK_NONE safety May 2024 Official leaderboard
7 Gemini 2.0 Pro 88.6% Feb 2025 Official leaderboard
8 GPT-4.5 88.3% Feb 2025 Official leaderboard
9 Gemini 2.5 Pro Mar 2025 88% Mar 2025 Official leaderboard
10 Gemini 2.0 Flash-Lite 87.9% Feb 2025 Official leaderboard
11 Gemini 1.5 Pro May 2024 87.8% BLOCK_NONE safety May 2024 Official leaderboard
12 o4-mini 87.5% High reasoning effort Apr 2025 Official leaderboard
13 Qwen2.5-VL-7B 87.2% Jan 2025 Official leaderboard
14 palmyra-vision-003 87.2% Official leaderboard
15 Qwen2.5-VL-72B 87% Jan 2025 Official leaderboard
16 o3 87% High reasoning effort Apr 2025 Official leaderboard
17 GPT-4o Nov 2024 86.8% Nov 2024 Official leaderboard
18 Gemini 1.0 Pro Vision 86.3% Jan 2024 Official leaderboard
19 GPT-4.1 85.7% Apr 2025 Official leaderboard
20 GPT-4.1 mini 85.6% Apr 2025 Official leaderboard
21 paligemma-3b-mix-224 85.3% Official leaderboard
22 o1 84.8% Dec 2024 Official leaderboard
23 Llama 4 Scout 84.5% Apr 2025 Official leaderboard
24 Llama 4 Maverick 84.2% Apr 2025 Official leaderboard
25 GPT-4o Aug 2024 83% Aug 2024 Official leaderboard
26 PaliGemma-3B-mix-448 82.3% May 2024 Official leaderboard
27 GPT-4o mini 82.2% Jul 2024 Official leaderboard
28 Qwen2.5-VL-32B 81.8% Mar 2025 Official leaderboard
29 GPT-4o May 2024 81.5% May 2024 Official leaderboard
30 GPT-4.1 nano 81.3% Apr 2025 Official leaderboard
31 GPT-4 Turbo Apr 2024 81.2% Apr 2024 Official leaderboard
32 Qwen2.5-VL-3B 80.8% Jan 2025 Official leaderboard
33 GPT-4 Turbo Nov 2023 79.6% Nov 2023 Official leaderboard
34 Pixtral 12B 63.5% Sep 2024 Official leaderboard
35 Claude 3.5 Sonnet (October 2024) 63% Oct 2024 Official leaderboard
36 Idefics2-8B 62.9% Apr 2024 Official leaderboard
37 Claude 3.5 Sonnet 62.1% Jun 2024 Official leaderboard
38 IDEFICS-9B-Instruct 60% Jul 2023 Official leaderboard
39 Claude 3 Sonnet 59.5% Feb 2024 Official leaderboard
40 Claude 3 Haiku 58.5% Mar 2024 Official leaderboard
41 IDEFICS-80B-Instruct 54% Jul 2023 Official leaderboard
42 Claude 3 Opus 52.5% Feb 2024 Official leaderboard
43 Qwen-VL-Chat 48.8% Aug 2023 Official leaderboard
#ModelScoreSettingDateSource
1 Qwen2-VL-72B-Instruct 91.2% Aug 2024 Official leaderboard
2 Qwen2-VL-7B-Instruct 90.2% Aug 2024 Official leaderboard
3 Gemini 2.0 Flash Feb 2025 89.9% Feb 2025 Official leaderboard
4 Gemini 1.5 Flash Sep 2024 89.4% Aug 2024 Official leaderboard
5 Gemini 1.5 Pro Sep 2024 89% Sep 2024 Official leaderboard
6 Gemini 1.5 Flash May 2024 88.8% BLOCK_NONE safety May 2024 Official leaderboard
7 Gemini 2.0 Pro 88.6% Feb 2025 Official leaderboard
8 GPT-4.5 88.3% Feb 2025 Official leaderboard
9 Gemini 2.5 Pro Mar 2025 88% Mar 2025 Official leaderboard
10 Gemini 2.0 Flash-Lite 87.9% Feb 2025 Official leaderboard
11 Gemini 1.5 Pro May 2024 87.8% BLOCK_NONE safety May 2024 Official leaderboard
12 o4-mini 87.5% High reasoning effort Apr 2025 Official leaderboard
13 Qwen2.5-VL-7B 87.2% Jan 2025 Official leaderboard
14 Qwen2.5-VL-72B 87% Jan 2025 Official leaderboard
15 o3 87% High reasoning effort Apr 2025 Official leaderboard
16 GPT-4o Nov 2024 86.8% Nov 2024 Official leaderboard
17 Gemini 1.0 Pro Vision 86.3% Jan 2024 Official leaderboard
18 GPT-4.1 85.7% Apr 2025 Official leaderboard
19 GPT-4.1 mini 85.6% Apr 2025 Official leaderboard
20 o1 84.8% Dec 2024 Official leaderboard
21 Llama 4 Scout 84.5% Apr 2025 Official leaderboard
22 Llama 4 Maverick 84.2% Apr 2025 Official leaderboard
23 GPT-4o Aug 2024 83% Aug 2024 Official leaderboard
24 PaliGemma-3B-mix-448 82.3% May 2024 Official leaderboard
25 GPT-4o mini 82.2% Jul 2024 Official leaderboard
26 Qwen2.5-VL-32B 81.8% Mar 2025 Official leaderboard
27 GPT-4o May 2024 81.5% May 2024 Official leaderboard
28 GPT-4.1 nano 81.3% Apr 2025 Official leaderboard
29 GPT-4 Turbo Apr 2024 81.2% Apr 2024 Official leaderboard
30 Qwen2.5-VL-3B 80.8% Jan 2025 Official leaderboard
31 GPT-4 Turbo Nov 2023 79.6% Nov 2023 Official leaderboard
32 Pixtral 12B 63.5% Sep 2024 Official leaderboard
33 Claude 3.5 Sonnet (October 2024) 63% Oct 2024 Official leaderboard
34 Idefics2-8B 62.9% Apr 2024 Official leaderboard
35 Claude 3.5 Sonnet 62.1% Jun 2024 Official leaderboard
36 IDEFICS-9B-Instruct 60% Jul 2023 Official leaderboard
37 Claude 3 Sonnet 59.5% Feb 2024 Official leaderboard
38 Claude 3 Haiku 58.5% Mar 2024 Official leaderboard
39 IDEFICS-80B-Instruct 54% Jul 2023 Official leaderboard
40 Claude 3 Opus 52.5% Feb 2024 Official leaderboard
41 Qwen-VL-Chat 48.8% Aug 2023 Official leaderboard

6 models have held number one on VHELM’s leaderboard since its release in November 2023

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

Hugging Face OpenAI Google DeepMind Other labs
VHELM leaderboard Passed 85% in Jan 2024, by Gemini 1.0 Pro Vision

Source: VHELM’s leaderboard

Gemini 1.0 Pro Vision on the official leaderboard passed 85% in January 2024

That was 2 months after its release. By the same age, 8% of the benchmarks from 2023 had passed 85%. All the scores we have, one dot each, so a model tested at several settings appears more than once.

Plot with other benchmarks
On VHELM’s leaderboard Best so far
0% 20% 40% 60% 80% 100% 2024 2025 Near the ceiling (85%) Qwen2-VL-72B-Instruct Gemini 1.0 Pro Vision GPT-4 Turbo IDEFICS-9B-Instruct 0% 20% 40% 60% 80% 100% 2024 2025 Near the ceiling (85%) Qwen2-VL-72B-Instruct GPT-4 Turbo

Source: VHELM’s leaderboard

No lab model card reports it yet

We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.

More in Vision

All 636 in Vision
  • MMMU In use 11.5K college-level multimodal questions across 30 subjects. Multiple-choice and open accuracy. Qwen 3.6 Plus 86%
  • MMMU-Pro In use Harder MMMU: college-level multimodal questions with more options and vision-only inputs to remove text shortcuts. GPT-5.6 Sol 84.6%
  • CharXiv In use Chart understanding on real arXiv figures. Reasoning questions require synthesizing information from complex charts. Claude Fable 5 93.5%
  • MathVista In use Math reasoning in visual contexts (charts, geometry, puzzles) from 31 datasets. Answer accuracy. Seed2.1 Pro 90.7%
  • MATH-Vision (MathVision) In use 3,040 competition math problems with visual context. Answer accuracy. Qwen 3.8 Max 97.7%
  • MMStar In use Models answer 1,500 human-curated image questions that cannot be solved without the image, across 6 core capabilities. Scored by accuracy. Qwen 3.8 Max 85.9%

Data audit

All 8 checks pass. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 2 pass 3 notes
  • Pass Independent runs Scores from the official leaderboard (43 entries).
  • Pass Models scored 41 models we track, plus 2 other entries on its leaderboard.
  • Note Newest score The newest came 19 months ago, in April 2025, for o3. That’s expected for a saturated test.
  • Note Frontier models None of the top 10 has a score, which is fine for a retired test. The strongest tested is o3 (#84).
  • Note Headroom Saturated. The best is 91.2%, past 85% since January 2024.
Leaderboard import 3 pass
  • Pass Importer We read the board on VHELM: 46 rows for models we track and 2 other entries at the last read.
  • Pass Last read November 20th, 2026.
  • Pass Read worked The last read worked.
Consistency 1 pass 1 note
  • Note Lab claims No model has both a lab number and an independent run.
  • Pass Known issues No published audit found problems.

How we check