PISA-Bench
Models answer illustrated PISA questions in six languages. Parallel questions let researchers compare language effects on reading, science and mathematical reasoning.
- Made by
- Patrick Haller and colleagues
- Released
- October 2025
- Kind
- General visual questions
- Level
- School
- Scored
- Checked automatically
- Answers
- Multiple choice
- Tasks
- All public
- Measures
- Capability
- Language
- Many languages
- Interaction
- Single turn
- Items from
- Exams and contests
- Refresh
- One fixed set
- Run by
- Anyone
- In views
- Vision, Many languages
- Strongest model tested
-
Gemma 3 27B #165 on the capability index
- Last model added
- October 2025
- Scores refreshed
- November 20th, 2026
GPT-4o (snapshot unspecified) tops the leaderboard with 67.8% Qwen2.5-VL-72B leads with 63.3% among models released by October 2025
Best score per entry from the benchmark's own leaderboard, read on November 20th, 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 best model we track is Qwen2.5-VL-72B, at 63.3%. This view shows only models we track, each at its best.
The newest model with a score here came out in October 2025, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 67.8 GPT-4o (snapshot unspecified)
- 63.3
Qwen2.5-VL-72B
- 61.7 Claude 3.5 Haiku
- 60.9
Gemma 3 27B
- 54.8
Qwen3-VL-8B-Instruct
- 52.6
Gemma 3 12B
- 50.9
Qwen2.5-VL-7B
- 50.3
Qwen3-VL-30B-A3B-Instruct
- 49.5
Qwen3-VL-4B-Instruct
- 45.2
Qwen2.5-VL-32B
- 42.5
Qwen2.5-VL-3B
- 40.9
Idefics3-8B-Llama3
- 38.7 llava-v1.6-34b-hf
- 38.3
SmolVLM2-256M-Video-Instruct
- 38.1
Gemma 3 4B
- 31.6 llava-1.5-7b-hf
- 31.5 llava-1.5-13b-hf
- 30.4
SmolVLM2-2.2B-Instruct
- 20.5
SmolVLM2-500M-Video-Instruct
- 1 GPT-4o (snapshot unspecified) 67.8%
- 2
Qwen2.5-VL-72B 63.3%
- 3 Claude 3.5 Haiku 61.7%
- 4
Gemma 3 27B 60.9%
- 5
Qwen3-VL-8B-Instruct 54.8%
- 6
Gemma 3 12B 52.6%
- 7
Qwen2.5-VL-7B 50.9%
- 8
Qwen3-VL-30B-A3B-Instruct 50.3%
- 9
Qwen3-VL-4B-Instruct 49.5%
- 10
Qwen2.5-VL-32B 45.2%
- 11
Qwen2.5-VL-3B 42.5%
- 12
Idefics3-8B-Llama3 40.9%
- 13 llava-v1.6-34b-hf 38.7%
- 14
SmolVLM2-256M-Video-Instruct 38.3%
- 15
Gemma 3 4B 38.1%
- 16 llava-1.5-7b-hf 31.6%
- 17 llava-1.5-13b-hf 31.5%
- 18
SmolVLM2-2.2B-Instruct 30.4%
- 19
SmolVLM2-500M-Video-Instruct 20.5%
Source: Official PISA-Bench leaderboard
- 63.3
Qwen2.5-VL-72B
- 61.7 Claude 3.5 Haiku
- 60.9
Gemma 3 27B
- 54.8
Qwen3-VL-8B-Instruct
- 52.6
Gemma 3 12B
- 50.9
Qwen2.5-VL-7B
- 50.3
Qwen3-VL-30B-A3B-Instruct
- 49.5
Qwen3-VL-4B-Instruct
- 45.2
Qwen2.5-VL-32B
- 42.5
Qwen2.5-VL-3B
- 40.9
Idefics3-8B-Llama3
- 38.3
SmolVLM2-256M-Video-Instruct
- 38.1
Gemma 3 4B
- 30.4
SmolVLM2-2.2B-Instruct
- 20.5
SmolVLM2-500M-Video-Instruct
- 1
Qwen2.5-VL-72B 63.3%
- 2 Claude 3.5 Haiku 61.7%
- 3
Gemma 3 27B 60.9%
- 4
Qwen3-VL-8B-Instruct 54.8%
- 5
Gemma 3 12B 52.6%
- 6
Qwen2.5-VL-7B 50.9%
- 7
Qwen3-VL-30B-A3B-Instruct 50.3%
- 8
Qwen3-VL-4B-Instruct 49.5%
- 9
Qwen2.5-VL-32B 45.2%
- 10
Qwen2.5-VL-3B 42.5%
- 11
Idefics3-8B-Llama3 40.9%
- 12
SmolVLM2-256M-Video-Instruct 38.3%
- 13
Gemma 3 4B 38.1%
- 14
SmolVLM2-2.2B-Instruct 30.4%
- 15
SmolVLM2-500M-Video-Instruct 20.5%
Source: Official PISA-Bench leaderboard
| # | Entry | Score | Setting | Date | Link | Source |
|---|---|---|---|---|---|---|
| 1 | GPT-4o (snapshot unspecified) | 67.8% | Official leaderboard | |||
| 2 | | 63.3% | Jan 2025 | huggingface.co | Official leaderboard | |
| 3 | Claude 3.5 Haiku | 61.7% | Oct 2024 | Official leaderboard | ||
| 4 | | 60.9% | Mar 2025 | huggingface.co | Official leaderboard | |
| 5 | | 54.8% | Oct 2025 | huggingface.co | Official leaderboard | |
| 6 | | 52.6% | Mar 2025 | huggingface.co | Official leaderboard | |
| 7 | | 50.9% | Jan 2025 | huggingface.co | Official leaderboard | |
| 8 | | 50.3% | Oct 2025 | huggingface.co | Official leaderboard | |
| 9 | | 49.5% | Oct 2025 | huggingface.co | Official leaderboard | |
| 10 | | 45.2% | Mar 2025 | huggingface.co | Official leaderboard | |
| 11 | | 42.5% | Jan 2025 | huggingface.co | Official leaderboard | |
| 12 | | 40.9% | Aug 2024 | huggingface.co | Official leaderboard | |
| 13 | llava-v1.6-34b-hf | 38.7% | huggingface.co | Official leaderboard | ||
| 14 | | 38.3% | Feb 2025 | huggingface.co | Official leaderboard | |
| 15 | | 38.1% | Mar 2025 | huggingface.co | Official leaderboard | |
| 16 | llava-1.5-7b-hf | 31.6% | huggingface.co | Official leaderboard | ||
| 17 | llava-1.5-13b-hf | 31.5% | huggingface.co | Official leaderboard | ||
| 18 | | 30.4% | Feb 2025 | huggingface.co | Official leaderboard | |
| 19 | | 20.5% | Feb 2025 | huggingface.co | Official leaderboard |
| # | Model | Score | Setting | Date | Link | Source |
|---|---|---|---|---|---|---|
| 1 | | 63.3% | Jan 2025 | huggingface.co | Official leaderboard | |
| 2 | Claude 3.5 Haiku | 61.7% | Oct 2024 | Official leaderboard | ||
| 3 | | 60.9% | Mar 2025 | huggingface.co | Official leaderboard | |
| 4 | | 54.8% | Oct 2025 | huggingface.co | Official leaderboard | |
| 5 | | 52.6% | Mar 2025 | huggingface.co | Official leaderboard | |
| 6 | | 50.9% | Jan 2025 | huggingface.co | Official leaderboard | |
| 7 | | 50.3% | Oct 2025 | huggingface.co | Official leaderboard | |
| 8 | | 49.5% | Oct 2025 | huggingface.co | Official leaderboard | |
| 9 | | 45.2% | Mar 2025 | huggingface.co | Official leaderboard | |
| 10 | | 42.5% | Jan 2025 | huggingface.co | Official leaderboard | |
| 11 | | 40.9% | Aug 2024 | huggingface.co | Official leaderboard | |
| 12 | | 38.3% | Feb 2025 | huggingface.co | Official leaderboard | |
| 13 | | 38.1% | Mar 2025 | huggingface.co | Official leaderboard | |
| 14 | | 30.4% | Feb 2025 | huggingface.co | Official leaderboard | |
| 15 | | 20.5% | Feb 2025 | huggingface.co | Official leaderboard |
Rankings by part
We read 6 parts from the benchmark’s own leaderboard, each ranking models on one part of the test, such as English, German and French. Qwen2.5-VL-72B leads 3 of them, and 2 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.
English
15 models we track have a score here.
- 1
Qwen2.5-VL-72B 69.4%
- 2 Claude 3.5 Haiku 62.9%
- 3
Gemma 3 27B 60.5%
- 4
Qwen3-VL-8B-Instruct 58.9%
- 5
Gemma 3 12B 58.1%
- 6
Qwen3-VL-30B-A3B-Instruct ↑1 57.3%
- 7
Qwen2.5-VL-7B ↓1 52.4%
- 8
Qwen2.5-VL-32B ↑1 51.6%
- 9
Qwen3-VL-4B-Instruct ↓1 48.4%
- 10
Idefics3-8B-Llama3 ↑1 47.6%
German
15 models we track have a score here.
- 1
Gemma 3 27B ↑2 63.9%
- 2
Qwen2.5-VL-72B ↓1 58.2%
- 3 Claude 3.5 Haiku ↓1 56.6%
- 4
Gemma 3 12B ↑1 52.5%
- 5
Qwen3-VL-8B-Instruct ↓1 52.5%
- 6
Qwen3-VL-4B-Instruct ↑2 50%
- 7
Qwen2.5-VL-7B ↓1 48.4%
- 8
Qwen3-VL-30B-A3B-Instruct ↓1 48.4%
- 9
Qwen2.5-VL-32B 44.3%
- 10
Idefics3-8B-Llama3 ↑1 42.6%
French
15 models we track have a score here.
- 1 Claude 3.5 Haiku ↑1 64.8%
- 2
Gemma 3 27B ↑1 61.5%
- 3
Qwen2.5-VL-72B ↓2 60.7%
- 4
Qwen3-VL-8B-Instruct 57.4%
- 5
Qwen2.5-VL-7B ↑1 56.6%
- 6
Qwen3-VL-4B-Instruct ↑2 51.6%
- 7
Gemma 3 12B ↓2 50.8%
- 8
Qwen3-VL-30B-A3B-Instruct ↓1 50.8%
- 9
Qwen2.5-VL-32B 46.7%
- 10
Qwen2.5-VL-3B 42.6%
Italian
15 models we track have a score here.
- 1
Qwen2.5-VL-72B 64.8%
- 2
Gemma 3 27B ↑1 63.9%
- 3 Claude 3.5 Haiku ↓1 59.8%
- 4
Qwen2.5-VL-7B ↑2 54.1%
- 5
Gemma 3 12B 51.6%
- 6
Qwen3-VL-30B-A3B-Instruct ↑1 50.8%
- 7
Qwen3-VL-4B-Instruct ↑1 49.2%
- 8
Qwen3-VL-8B-Instruct ↓4 48.4%
- 9
Qwen2.5-VL-32B 45.1%
- 10
Qwen2.5-VL-3B 43.4%
Spanish
15 models we track have a score here.
- 1
Qwen2.5-VL-72B 63.1%
- 2 Claude 3.5 Haiku 61.5%
- 3
Gemma 3 27B 61.5%
- 4
Qwen3-VL-8B-Instruct 55.7%
- 5
Gemma 3 12B 48.4%
- 6
Qwen2.5-VL-7B 46.7%
- 7
Qwen3-VL-4B-Instruct ↑1 45.1%
- 8
SmolVLM2-256M-Video-Instruct ↑4 44.3%
- 9
Qwen3-VL-30B-A3B-Instruct ↓2 44.3%
- 10
Idefics3-8B-Llama3 ↑1 42.6%
Chinese
15 models we track have a score here.
- 1 Claude 3.5 Haiku ↑1 64.8%
- 2
Qwen2.5-VL-72B ↓1 63.9%
- 3
Qwen3-VL-8B-Instruct ↑1 55.7%
- 4
Gemma 3 27B ↓1 54.1%
- 5
Gemma 3 12B 54.1%
- 6
Qwen3-VL-4B-Instruct ↑2 52.5%
- 7
Qwen3-VL-30B-A3B-Instruct 50%
- 8
Qwen2.5-VL-7B ↓2 47.5%
- 9
Qwen2.5-VL-32B 44.3%
- 10
Qwen2.5-VL-3B 40.2%
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official PISA-Bench leaderboard
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official PISA-Bench 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
- 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
9 of 10 checks pass, 1 warning. Look first at the Newest score check. The newest came 13 months ago, in October 2025, for Qwen3-VL-4B-Instruct. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 1 warning 3 pass 1 note
- Pass Independent runs Scores from the official leaderboard (19 entries).
- Pass Models scored 15 models we track, plus 4 other entries on its leaderboard.
- Warning Newest score The newest came 13 months ago, in October 2025, for Qwen3-VL-4B-Instruct.
- Note Frontier models None of the top 10 has a score. The strongest tested is Gemma 3 27B (#165).
- Pass Headroom The best is 63.3%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on huggingface.co: 15 rows for models we track and 4 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.