Benchmark Almanac

WorldVQA

Qwen3.7-Plus 61.1% leads since June 2026
Gemini 3 Pro 47.4% leads since January 2026
Qwen3.7-Plus 61.1% leads since June 2026
In use

Model names visual entities in 3,500 image-question pairs across nine categories, isolating memorized visual world knowledge. Exact-answer accuracy.

Lab-created public benchmark (Moonshot AI). Kimi K3 reports a 'ForceAnswer' variant.

Made by
Moonshot AI
Released
January 2026
Kind
General visual questions
Level
University
Scored
Graded by a model
Answers
Short answer
Tasks
All public
Strongest model tested
Kimi K3 #24 on the capability index
Last in a launch post
October 2026
Last model added
October 2026
Scores refreshed
November 12th, 2026

In the authors’ words

We introduce WorldVQA, a benchmark designed to evaluate the atomic visual world knowledge of Multimodal Large Language Models (MLLMs). Unlike current evaluations, which often conflate visual knowledge retrieval with reasoning, WorldVQA decouples these capabilities to strictly measure "what the model memorizes." The benchmark assesses the atomic capability of grounding and naming visual entities across a stratified taxonomy, spanning from common head-class objects to long-tail rarities. We expect WorldVQA to serve as a rigorous test for visual factuality, thereby establishing a standard for assessing the encyclopedic breadth and hallucination rates of current and next-generation frontier models.

Abstract of WorldVQA: Measuring Atomic World Knowledge in Multimodal Large Language Models (arXiv:2602.02537, CC0)
9
Models scored in lab cards
8
Launch posts list it, from 5 labs
4
System cards report it, from 2 labs

Qwen3.7-Plus leads with 61.1%

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.

Gemini 3 Pro leads with 47.4% across all independent runs

Best score per model across the official leaderboard, Epoch AI’s own runs and any board Epoch AI copied. Each bar and row names its source.

Qwen3.7-Plus leads with 61.1%, claims included

Best score per model from every source, including labs’ own numbers and claims Epoch AI collected from reports and papers. Labs pick their own settings, so a claim can beat an independent run of the same model.

  1. 61.1 Qwen3.7-Plus
  2. 58.1 Step 3.7 Flash
  3. 53.2 Qwen 3.8 Max
  4. 53.0 Seed2.1 Pro
  5. 51.0 Kimi K3
  6. 50.2 Seed2.0 Lite (0428)
  7. 49.9 Seed2.0 Pro
  8. 49.4 Muse Spark Open
  9. 46.3 Kimi K2.5
  1. 1 Qwen3.7-Plus 61.1%
  2. 2 Step 3.7 Flash 58.1%
  3. 3 Qwen 3.8 Max 53.2%
  4. 4 Seed2.1 Pro 53%
  5. 5 Kimi K3 51%
  6. 6 Seed2.0 Lite (0428) 50.2%
  7. 7 Seed2.0 Pro 49.9%
  8. 8 Muse Spark Open 49.4%
  9. 9 Kimi K2.5 46.3%

Source: Labs’ model cards

22 models. Scroll sideways for the rest.

  1. 47.4 Gemini 3 Pro
  2. 46.3 Kimi K2.5
  3. 36.8 Claude Opus 4.5
  4. 28.0 GPT-5.2
  5. 24.5 GPT-5.1
  6. 23.5 Qwen3-VL-235B-A22B-Instruct
  7. 22.3 Qwen3-VL-235B-A22B
  8. 21.1 Grok 4.1 Fast
  9. 20.0 Claude Sonnet 4.5
  10. 19.0 GLM-4.6V
  11. 18.9 Grok 4 Fast
  12. 18.9 Qwen3-VL-30B-A3B-Instruct
  13. 17.7 Qwen3-VL-32B-Instruct
  14. 17.6 Qwen3-VL-30B-A3B-Thinking
  15. 16.4 Qwen3-VL-32B-Thinking
  16. 16.1 Qwen3-VL-8B-Instruct
  17. 14.9 Qwen3-VL-4B-Instruct
  18. 14.8 GLM-4.6V-Flash
  19. 11.9 Qwen3-VL-8B-Thinking
  20. 10.9 Qwen3-VL-4B-Thinking
  21. 9.80 Qwen3-VL-2B-Instruct
  22. 7.10 Qwen3-VL-2B-Thinking
  1. 1 Gemini 3 Pro 47.4%
  2. 2 Kimi K2.5 46.3%
  3. 3 Claude Opus 4.5 36.8%
  4. 4 GPT-5.2 28%
  5. 5 GPT-5.1 24.5%
  6. 6 Qwen3-VL-235B-A22B-Instruct 23.5%
  7. 7 Qwen3-VL-235B-A22B thinking 22.3%
  8. 8 Grok 4.1 Fast reasoning 21.1%
  9. 9 Claude Sonnet 4.5 20%
  10. 10 GLM-4.6V 19%
  11. 11 Grok 4 Fast reasoning 18.9%
  12. 12 Qwen3-VL-30B-A3B-Instruct 18.9%
  13. 13 Qwen3-VL-32B-Instruct 17.7%
  14. 14 Qwen3-VL-30B-A3B-Thinking 17.6%
  15. 15 Qwen3-VL-32B-Thinking 16.4%
  16. 16 Qwen3-VL-8B-Instruct 16.1%
  17. 17 Qwen3-VL-4B-Instruct 14.9%
  18. 18 GLM-4.6V-Flash 14.8%
  19. 19 Qwen3-VL-8B-Thinking 11.9%
  20. 20 Qwen3-VL-4B-Thinking 10.9%
  21. 21 Qwen3-VL-2B-Instruct 9.80%
  22. 22 Qwen3-VL-2B-Thinking 7.10%

Source: Official WorldVQA leaderboard

30 models. Scroll sideways for the rest.

  1. 61.1 Qwen3.7-Plus
  2. 58.1 Step 3.7 Flash
  3. 53.2 Qwen 3.8 Max
  4. 53.0 Seed2.1 Pro
  5. 51.0 Kimi K3
  6. 50.2 Seed2.0 Lite (0428)
  7. 49.9 Seed2.0 Pro
  8. 49.4 Muse Spark Open
  9. 47.4 Gemini 3 Pro
  10. 46.3 Kimi K2.5
  11. 36.8 Claude Opus 4.5
  12. 28.0 GPT-5.2
  13. 24.5 GPT-5.1
  14. 23.5 Qwen3-VL-235B-A22B-Instruct
  15. 22.3 Qwen3-VL-235B-A22B
  16. 21.1 Grok 4.1 Fast
  17. 20.0 Claude Sonnet 4.5
  18. 19.0 GLM-4.6V
  19. 18.9 Grok 4 Fast
  20. 18.9 Qwen3-VL-30B-A3B-Instruct
  21. 17.7 Qwen3-VL-32B-Instruct
  22. 17.6 Qwen3-VL-30B-A3B-Thinking
  23. 16.4 Qwen3-VL-32B-Thinking
  24. 16.1 Qwen3-VL-8B-Instruct
  25. 14.9 Qwen3-VL-4B-Instruct
  26. 14.8 GLM-4.6V-Flash
  27. 11.9 Qwen3-VL-8B-Thinking
  28. 10.9 Qwen3-VL-4B-Thinking
  29. 9.80 Qwen3-VL-2B-Instruct
  30. 7.10 Qwen3-VL-2B-Thinking
  1. 1 Qwen3.7-Plus 61.1%
  2. 2 Step 3.7 Flash 58.1%
  3. 3 Qwen 3.8 Max 53.2%
  4. 4 Seed2.1 Pro 53%
  5. 5 Kimi K3 51%
  6. 6 Seed2.0 Lite (0428) 50.2%
  7. 7 Seed2.0 Pro 49.9%
  8. 8 Muse Spark Open 49.4%
  9. 9 Gemini 3 Pro 47.4%
  10. 10 Kimi K2.5 46.3%
  11. 11 Claude Opus 4.5 36.8%
  12. 12 GPT-5.2 28%
  13. 13 GPT-5.1 24.5%
  14. 14 Qwen3-VL-235B-A22B-Instruct 23.5%
  15. 15 Qwen3-VL-235B-A22B thinking 22.3%
  16. 16 Grok 4.1 Fast reasoning 21.1%
  17. 17 Claude Sonnet 4.5 20%
  18. 18 GLM-4.6V 19%
  19. 19 Grok 4 Fast reasoning 18.9%
  20. 20 Qwen3-VL-30B-A3B-Instruct 18.9%
  21. 21 Qwen3-VL-32B-Instruct 17.7%
  22. 22 Qwen3-VL-30B-A3B-Thinking 17.6%
  23. 23 Qwen3-VL-32B-Thinking 16.4%
  24. 24 Qwen3-VL-8B-Instruct 16.1%
  25. 25 Qwen3-VL-4B-Instruct 14.9%
  26. 26 GLM-4.6V-Flash 14.8%
  27. 27 Qwen3-VL-8B-Thinking 11.9%
  28. 28 Qwen3-VL-4B-Thinking 10.9%
  29. 29 Qwen3-VL-2B-Instruct 9.80%
  30. 30 Qwen3-VL-2B-Thinking 7.10%

Sources: Labs’ model cards and official WorldVQA leaderboard

#ModelScoreSettingDateSource
1 Qwen3.7-Plus 61.1% Jun 2026 Lab model card
2 Step 3.7 Flash 58.1% May 2026 Lab model card
3 Qwen 3.8 Max 53.2% Aug 2026 Lab model card
4 Seed2.1 Pro 53% Jun 2026 Lab model card
5 Kimi K3 51% Jul 2026 Lab model card
6 Seed2.0 Lite (0428) 50.2% Apr 2026 Lab model card
7 Seed2.0 Pro 49.9% Feb 2026 Lab model card
8 Muse Spark Open 49.4% Oct 2026 Lab model card
9 Kimi K2.5 46.3% Jan 2026 Lab model card
#ModelScoreSettingDateSource
1 Gemini 3 Pro 47.4% Nov 2025 Official leaderboard
2 Kimi K2.5 46.3% Jan 2026 Official leaderboard
3 Claude Opus 4.5 36.8% Nov 2025 Official leaderboard
4 GPT-5.2 28% Dec 2025 Official leaderboard
5 GPT-5.1 24.5% Nov 2025 Official leaderboard
6 Qwen3-VL-235B-A22B-Instruct 23.5% Sep 2025 Official leaderboard
7 Qwen3-VL-235B-A22B 22.3% thinking Sep 2025 Official leaderboard
8 Grok 4.1 Fast 21.1% reasoning Nov 2025 Official leaderboard
9 Claude Sonnet 4.5 20% Sep 2025 Official leaderboard
10 GLM-4.6V 19% Dec 2025 Official leaderboard
11 Grok 4 Fast 18.9% reasoning Sep 2025 Official leaderboard
12 Qwen3-VL-30B-A3B-Instruct 18.9% Oct 2025 Official leaderboard
13 Qwen3-VL-32B-Instruct 17.7% Oct 2025 Official leaderboard
14 Qwen3-VL-30B-A3B-Thinking 17.6% Oct 2025 Official leaderboard
15 Qwen3-VL-32B-Thinking 16.4% Oct 2025 Official leaderboard
16 Qwen3-VL-8B-Instruct 16.1% Oct 2025 Official leaderboard
17 Qwen3-VL-4B-Instruct 14.9% Oct 2025 Official leaderboard
18 GLM-4.6V-Flash 14.8% Dec 2025 Official leaderboard
19 Qwen3-VL-8B-Thinking 11.9% Oct 2025 Official leaderboard
20 Qwen3-VL-4B-Thinking 10.9% Oct 2025 Official leaderboard
21 Qwen3-VL-2B-Instruct 9.80% Oct 2025 Official leaderboard
22 Qwen3-VL-2B-Thinking 7.10% Oct 2025 Official leaderboard
#ModelScoreSettingDateSource
1 Qwen3.7-Plus 61.1% Jun 2026 Lab model card
2 Step 3.7 Flash 58.1% May 2026 Lab model card
3 Qwen 3.8 Max 53.2% Aug 2026 Lab model card
4 Seed2.1 Pro 53% Jun 2026 Lab model card
5 Kimi K3 51% Jul 2026 Lab model card
6 Seed2.0 Lite (0428) 50.2% Apr 2026 Lab model card
7 Seed2.0 Pro 49.9% Feb 2026 Lab model card
8 Muse Spark Open 49.4% Oct 2026 Lab model card
9 Gemini 3 Pro 47.4% Nov 2025 Official leaderboard
10 Kimi K2.5 46.3% Jan 2026 Official leaderboard
11 Claude Opus 4.5 36.8% Nov 2025 Official leaderboard
12 GPT-5.2 28% Dec 2025 Official leaderboard
13 GPT-5.1 24.5% Nov 2025 Official leaderboard
14 Qwen3-VL-235B-A22B-Instruct 23.5% Sep 2025 Official leaderboard
15 Qwen3-VL-235B-A22B 22.3% thinking Sep 2025 Official leaderboard
16 Grok 4.1 Fast 21.1% reasoning Nov 2025 Official leaderboard
17 Claude Sonnet 4.5 20% Sep 2025 Official leaderboard
18 GLM-4.6V 19% Dec 2025 Official leaderboard
19 Grok 4 Fast 18.9% reasoning Sep 2025 Official leaderboard
20 Qwen3-VL-30B-A3B-Instruct 18.9% Oct 2025 Official leaderboard
21 Qwen3-VL-32B-Instruct 17.7% Oct 2025 Official leaderboard
22 Qwen3-VL-30B-A3B-Thinking 17.6% Oct 2025 Official leaderboard
23 Qwen3-VL-32B-Thinking 16.4% Oct 2025 Official leaderboard
24 Qwen3-VL-8B-Instruct 16.1% Oct 2025 Official leaderboard
25 Qwen3-VL-4B-Instruct 14.9% Oct 2025 Official leaderboard
26 GLM-4.6V-Flash 14.8% Dec 2025 Official leaderboard
27 Qwen3-VL-8B-Thinking 11.9% Oct 2025 Official leaderboard
28 Qwen3-VL-4B-Thinking 10.9% Oct 2025 Official leaderboard
29 Qwen3-VL-2B-Instruct 9.80% Oct 2025 Official leaderboard
30 Qwen3-VL-2B-Thinking 7.10% Oct 2025 Official leaderboard

The benchmark's own leaderboard has 26 entries, 22 of them models we track. Switch the scores above to rank them together.

5 models have held number one in labs’ own numbers since its release in January 2026 Number one over time 5 models have held number one with claims included since its release in January 2026

Lab claims count from the day the card came out, independent runs from the model’s release. We draw one row each for the benchmark’s own leaderboard and what labs claim in their cards.

Google DeepMind Moonshot AI ByteDance Seed StepFun Qwen
Own leaderboard Not tested since Jan 2026
Gemini 3 Pro 47.4%
Lab claims 5 leaders
Qwen3.7-Plus 61.1%

Sources: Official WorldVQA leaderboard and labs’ model cards

Scores over time

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

On the benchmark’s own leaderboard Reported by the lab in its model card Best so far
0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) Qwen3.7-Plus Seed2.0 Lite (0428) Gemini 3 Pro GPT-5.1 Grok 4 Fast 0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) Qwen3.7-Plus Seed2.0 Lite (0428)

Sources: Official WorldVQA leaderboard and labs’ model cards

4 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.

Moonshot AI
ByteDance Seed
StepFun
Qwen
Meta
DateLaunch postScoreAs the post writes it
Oct 15, 2026 Muse Spark Open 49.4 WorldVQA (w/ tools)
Aug 3, 2026 Qwen 3.8 Max 53.2 WorldVQA
Jul 16, 2026 Kimi K3 51 WorldVQA ForceAnswer
Jun 23, 2026 Seed2.1 Pro 53 WorldVQA
Jun 1, 2026 Qwen3.7-Plus 61.1 WorldVQA (with search)
May 29, 2026 Step 3.7 Flash 58.1 WorldVQA
Apr 28, 2026 Seed2.0 Lite (0428) 50.2 WorldVQA
Jan 27, 2026 Kimi K2.5 46.3 WorldVQA

4 system cards report it, from 2 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
Jul 27, 2026 Kimi K3 technical report 51 WorldVQA ForceAnswer (Table 2) Capabilities
Jun 23, 2026 Seed2.1 Pro model card 53 WorldVQA (Table 9) Capabilities
Feb 14, 2026 Seed2.0 Pro model card 49.9 WorldVQA (Table 8) Capabilities
Jan 30, 2026 Kimi K2.5 technical report 46.3 WorldVQA (Table 4) Capabilities

One lab claim next to the leaderboard

Each score a lab reported for its own model, beside the same model's best score on the leaderboard. Gaps mostly come from different harnesses, effort levels and task subsets. 1 of them match, which usually means one side copied the other. Every benchmark.

Model Lab says Board says Gap
Kimi K2.5 46.3% WorldVQA 46.3% Leaderboard Same

More in Vision

All 281 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. 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%
  • 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%
  • 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%