Benchmark Almanac

Math-VR

Qwen3-VL-235B-A22B 66.8% leads since October 2025
Only in system cards

Models answer 5,000 English and Chinese math questions whose solutions need visual reasoning, such as drawing auxiliary lines or plotting functions. Answers are checked against ground truth.

The benchmark reports Answer Correctness and a partial-credit Process Score. Its 178K-sample training set ships separately. Qwen reports it in the Qwen3-VL report.

Made by
HKU MMLab, CUHK and Meituan
Released
October 2025
Kind
Visual math and diagrams
Level
School
Scored
Checked automatically
Answers
Short answer
Tasks
All public
Strongest model tested
GPT-5 #70 on the capability index
Last in a system card
November 2025
Last model added
November 2025
Scores refreshed
November 12th, 2026
28
Entries on its own leaderboard, 26 models we track
1
System card reports it, from 1 lab

Qwen3-VL-235B-A22B leads with 66.8% among models released by September 2025

Best score per entry from the benchmark's own leaderboard, read on November 12th, 2026. 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.

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

28 entries. Scroll sideways for the rest.

  1. 66.8 Qwen3-VL-235B-A22B
  2. 65.0 Qwen3-VL-235B-A22B-Instruct
  3. 64.7 Gemini 2.5 Pro
  4. 60.5 Gemini 2.5 Flash
  5. 59.3 o3
  6. 58.4 Seed1.6-Thinking
  7. 58.1 GPT-5
  8. 54.3 Claude Opus 4.1
  9. 53.4 Gemini 2.5 Flash Image Preview (nano-banana)
  10. 49.6 GLM-4.5V
  11. 48.3 MiMo-VL-7B-RL
  12. 40.8 InternVL3.5-8B
  13. 33.3 GPT-4.1 mini
  14. 29.0 GLM-4.1V-9B-Thinking
  15. 28.1 Claude Sonnet 4
  16. 26.0 GPT-4.1
  17. 22.1 CodePlot-CoT
  18. 20.6 Gemini 2.0 Flash
  19. 17.3 Keye-VL-1.5-8B
  20. 16.1 Gemma 3 27B
  21. 13.7 Qwen2.5-VL-72B
  22. 10.1 Bagel-Zebra-CoT
  23. 10.0 Qwen2.5-VL-32B
  24. 9.10 GPT-4.1 nano
  25. 7.60 BAGEL
  26. 5.30 Qwen2.5-VL-3B
  27. 4.30 GPT-4o
  28. 3.00 Qwen2.5-VL-7B
  1. 1 Qwen3-VL-235B-A22B thinking 66.8%
  2. 2 Qwen3-VL-235B-A22B-Instruct 65%
  3. 3 Gemini 2.5 Pro Jun 2025 64.7%
  4. 4 Gemini 2.5 Flash Jun 2025 60.5%
  5. 5 o3 59.3%
  6. 6 Seed1.6-Thinking 58.4%
  7. 7 GPT-5 thinking 58.1%
  8. 8 Claude Opus 4.1 54.3%
  9. 9 Gemini 2.5 Flash Image Preview (nano-banana) 53.4%
  10. 10 GLM-4.5V 49.6%
  11. 11 MiMo-VL-7B-RL 48.3%
  12. 12 InternVL3.5-8B 40.8%
  13. 13 GPT-4.1 mini 33.3%
  14. 14 GLM-4.1V-9B-Thinking 29%
  15. 15 Claude Sonnet 4 28.1%
  16. 16 GPT-4.1 26%
  17. 17 CodePlot-CoT 22.1%
  18. 18 Gemini 2.0 Flash Feb 2025 20.6%
  19. 19 Keye-VL-1.5-8B 17.3%
  20. 20 Gemma 3 27B 16.1%
  21. 21 Qwen2.5-VL-72B 13.7%
  22. 22 Bagel-Zebra-CoT 10.1%
  23. 23 Qwen2.5-VL-32B 10%
  24. 24 GPT-4.1 nano 9.10%
  25. 25 BAGEL 7.60%
  26. 26 Qwen2.5-VL-3B 5.30%
  27. 27 GPT-4o Nov 2024 4.30%
  28. 28 Qwen2.5-VL-7B 3.00%

Source: Official Math-VR leaderboard

#EntryScoreSettingDateSource
1 Qwen3-VL-235B-A22B 66.8% thinking Sep 2025 Official leaderboard
2 Qwen3-VL-235B-A22B-Instruct 65% Sep 2025 Official leaderboard
3 Gemini 2.5 Pro Jun 2025 64.7% Jun 2025 Official leaderboard
4 Gemini 2.5 Flash Jun 2025 60.5% Jun 2025 Official leaderboard
5 o3 59.3% Apr 2025 Official leaderboard
6 Seed1.6-Thinking 58.4% Jun 2025 Official leaderboard
7 GPT-5 58.1% thinking Aug 2025 Official leaderboard
8 Claude Opus 4.1 54.3% Aug 2025 Official leaderboard
9 Gemini 2.5 Flash Image Preview (nano-banana) 53.4% Official leaderboard
10 GLM-4.5V 49.6% Aug 2025 Official leaderboard
11 MiMo-VL-7B-RL 48.3% May 2025 Official leaderboard
12 InternVL3.5-8B 40.8% Aug 2025 Official leaderboard
13 GPT-4.1 mini 33.3% Apr 2025 Official leaderboard
14 GLM-4.1V-9B-Thinking 29% Jul 2025 Official leaderboard
15 Claude Sonnet 4 28.1% May 2025 Official leaderboard
16 GPT-4.1 26% Apr 2025 Official leaderboard
17 CodePlot-CoT 22.1% Official leaderboard
18 Gemini 2.0 Flash Feb 2025 20.6% Feb 2025 Official leaderboard
19 Keye-VL-1.5-8B 17.3% Aug 2025 Official leaderboard
20 Gemma 3 27B 16.1% Mar 2025 Official leaderboard
21 Qwen2.5-VL-72B 13.7% Jan 2025 Official leaderboard
22 Bagel-Zebra-CoT 10.1% Official leaderboard
23 Qwen2.5-VL-32B 10% Mar 2025 Official leaderboard
24 GPT-4.1 nano 9.10% Apr 2025 Official leaderboard
25 BAGEL 7.60% Official leaderboard
26 Qwen2.5-VL-3B 5.30% Jan 2025 Official leaderboard
27 GPT-4o Nov 2024 4.30% Nov 2024 Official leaderboard
28 Qwen2.5-VL-7B 3.00% Jan 2025 Official leaderboard

Labs report scores for 1 model in their cards. We plot every source in the charts below.

Number one over time

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

Qwen
Own leaderboard Not tested since Oct 2025
Qwen3-VL-235B-A22B 66.8%

Source: Official Math-VR leaderboard

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% 2025 2026 Near the ceiling (85%) Qwen3-VL-235B-A22B Gemini 2.5 Pro o3 GPT-4.1 mini Gemini 2.0 Flash Qwen2.5-VL-72B GPT-4o 0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) Qwen3-VL-235B-A22B o3

Sources: Official Math-VR leaderboard and labs’ model cards

Labs report it only in system cards

We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark. Some system cards do, below.

One system card reports it

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
Nov 26, 2025 Qwen3-VL-235B-A22B technical report 66.8 Math-VR (Table 2) 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
Qwen3-VL-235B-A22B 66.8% Math-VR (Table 2) 66.8% Leaderboard, thinking 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%