MatCha
Models answer multiple-choice questions about material structure, shape and properties using images from materials characterization instruments.
The authors provide 1,500 questions across 21 tasks. The ATOMIC team independently compares models on the transmission electron microscopy subset.
- Released
- September 2025
- Kind
- Lab and data work
- Level
- Expert
- Scored
- Checked automatically
- Answers
- Multiple choice
- Tasks
- All public
- Measures
- Capability
- Subject
- Materials engineering
- Language
- English
- Interaction
- Single turn
- Items from
- Real work and use
- Refresh
- One fixed set
- Run by
- Anyone
- Occupations
- Life, physical and social science
- In views
- Vision, Chemistry and materials
- Strongest model tested
-
Llama 4 Maverick #153 on the capability index
- Last model added
- April 2025
- Scores refreshed
- November 20th, 2026
Claude-3.5-Sonnet (snapshot unspecified) tops the leaderboard with 68.6% Llama 4 Maverick leads with 68.1% among models released by April 2025
Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart. 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 Llama 4 Maverick, at 68.1%. 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:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 68.6 Claude-3.5-Sonnet (snapshot unspecified)
- 68.1
Llama 4 Maverick
- 68.0 GPT-4o (snapshot unspecified)
- 57.1 Gemini-1.5-Pro (snapshot unspecified)
- 52.3 Gemini-1.5-Flash (snapshot unspecified)
- 52.2
InternVL3-38B
- 50.1 GPT-4o-mini (snapshot unspecified)
- 50.1
Qwen2.5-VL-32B
- 44.6
InternVL3-8B
- 42.7
Qwen2.5-VL-7B
- 40.3
Janus-Pro-7B
- 35.1
Gemma 3 4B
- 34.8 Llama-3.2-11B-Vision
- 29.7
LLaVA-v1.5-13B
- 24.1 llava-v1.5-7b
- 1 Claude-3.5-Sonnet (snapshot unspecified) 8-shot 68.6%
- 2
Llama 4 Maverick 16-shot 68.1%
- 3 GPT-4o (snapshot unspecified) 16-shot 68%
- 4 Gemini-1.5-Pro (snapshot unspecified) 4-shot 57.1%
- 5 Gemini-1.5-Flash (snapshot unspecified) 16-shot 52.3%
- 6
InternVL3-38B 2-shot 52.2%
- 7 GPT-4o-mini (snapshot unspecified) 16-shot 50.1%
- 8
Qwen2.5-VL-32B Chain of thought 50.1%
- 9
InternVL3-8B 8-shot 44.6%
- 10
Qwen2.5-VL-7B Chain of thought 42.7%
- 11
Janus-Pro-7B Chain of thought 40.3%
- 12
Gemma 3 4B 8-shot 35.1%
- 13 Llama-3.2-11B-Vision Chain of thought 34.8%
- 14
LLaVA-v1.5-13B 2-shot 29.7%
- 15 llava-v1.5-7b 2-shot 24.1%
Source: The MatCha paper
- 1
Llama 4 Maverick 16-shot 68.1%
- 2
InternVL3-38B 2-shot 52.2%
- 3
Qwen2.5-VL-32B Chain of thought 50.1%
- 4
InternVL3-8B 8-shot 44.6%
- 5
Qwen2.5-VL-7B Chain of thought 42.7%
- 6
Janus-Pro-7B Chain of thought 40.3%
- 7
Gemma 3 4B 8-shot 35.1%
- 8
LLaVA-v1.5-13B 2-shot 29.7%
- 9 llava-v1.5-7b 2-shot 24.1%
Source: The MatCha paper
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude-3.5-Sonnet (snapshot unspecified) | 68.6% | 8-shot | Official leaderboard | |
| 2 | | 68.1% | 16-shot | Apr 2025 | Official leaderboard |
| 3 | GPT-4o (snapshot unspecified) | 68% | 16-shot | Official leaderboard | |
| 4 | Gemini-1.5-Pro (snapshot unspecified) | 57.1% | 4-shot | Official leaderboard | |
| 5 | Gemini-1.5-Flash (snapshot unspecified) | 52.3% | 16-shot | Official leaderboard | |
| 6 | | 52.2% | 2-shot | Apr 2025 | Official leaderboard |
| 7 | GPT-4o-mini (snapshot unspecified) | 50.1% | 16-shot | Official leaderboard | |
| 8 | | 50.1% | Chain of thought | Mar 2025 | Official leaderboard |
| 9 | | 44.6% | 8-shot | Apr 2025 | Official leaderboard |
| 10 | | 42.7% | Chain of thought | Jan 2025 | Official leaderboard |
| 11 | | 40.3% | Chain of thought | Jan 2025 | Official leaderboard |
| 12 | | 35.1% | 8-shot | Mar 2025 | Official leaderboard |
| 13 | Llama-3.2-11B-Vision | 34.8% | Chain of thought | Official leaderboard | |
| 14 | | 29.7% | 2-shot | Oct 2023 | Official leaderboard |
| 15 | llava-v1.5-7b | 24.1% | 2-shot | Oct 2023 | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 68.1% | 16-shot | Apr 2025 | Official leaderboard |
| 2 | | 52.2% | 2-shot | Apr 2025 | Official leaderboard |
| 3 | | 50.1% | Chain of thought | Mar 2025 | Official leaderboard |
| 4 | | 44.6% | 8-shot | Apr 2025 | Official leaderboard |
| 5 | | 42.7% | Chain of thought | Jan 2025 | Official leaderboard |
| 6 | | 40.3% | Chain of thought | Jan 2025 | Official leaderboard |
| 7 | | 35.1% | 8-shot | Mar 2025 | Official leaderboard |
| 8 | | 29.7% | 2-shot | Oct 2023 | Official leaderboard |
| 9 | llava-v1.5-7b | 24.1% | 2-shot | Oct 2023 | Official leaderboard |
Rankings by part
We read 9 parts from the benchmark’s paper, each ranking models on one part of the test, such as Processing correlation, Morphology analysis and Structure analysis. Llama 4 Maverick leads every one of them. Arrows show how far a model moved from its overall rank among the models we track.
Processing correlation
9 models we track have a score here.
- 1
Llama 4 Maverick 88.2%
- 2
InternVL3-38B 75.2%
- 3
Qwen2.5-VL-7B ↑2 71.2%
- 4
Qwen2.5-VL-32B ↓1 69.3%
- 5
Gemma 3 4B ↑2 64.7%
- 6
InternVL3-8B ↓2 62.1%
- 7
Janus-Pro-7B ↓1 55.6%
- 8
LLaVA-v1.5-13B 43.1%
- 9 llava-v1.5-7b 26.1%
Morphology analysis
9 models we track have a score here.
- 1
Llama 4 Maverick 75.8%
- 2
InternVL3-38B 69.9%
- 3
Qwen2.5-VL-32B 67.8%
- 4
Qwen2.5-VL-7B ↑1 62.6%
- 5
InternVL3-8B ↓1 61.9%
- 6
Janus-Pro-7B 55.7%
- 7
Gemma 3 4B 50.9%
- 8
LLaVA-v1.5-13B 48.4%
- 9 llava-v1.5-7b 36%
Structure analysis
9 models we track have a score here.
- 1
Llama 4 Maverick 72.7%
- 2
Qwen2.5-VL-32B ↑1 63%
- 3
Janus-Pro-7B ↑3 61.4%
- 4
InternVL3-38B ↓2 61.1%
- 5
Qwen2.5-VL-7B 54%
- 6
InternVL3-8B ↓2 53%
- 7
Gemma 3 4B 49.5%
- 8
LLaVA-v1.5-13B 42.7%
- 9 llava-v1.5-7b 34.3%
Property analysis
9 models we track have a score here.
- 1
Llama 4 Maverick 68.7%
- 2
Qwen2.5-VL-32B ↑1 66.5%
- 3
InternVL3-38B ↓1 61.5%
- 4
Qwen2.5-VL-7B ↑1 59.3%
- 5
Janus-Pro-7B ↑1 57.7%
- 6
InternVL3-8B ↓2 56%
- 7
Gemma 3 4B 46.7%
- 8
LLaVA-v1.5-13B 43.4%
- 9 llava-v1.5-7b 31.3%
Generated VQA
9 models we track have a score here.
- 1
Llama 4 Maverick 64.6%
- 2
InternVL3-38B 53.2%
- 3
Qwen2.5-VL-32B 52.6%
- 4
Janus-Pro-7B ↑2 46.5%
- 5
Qwen2.5-VL-7B 46.1%
- 6
InternVL3-8B ↓2 45.4%
- 7
Gemma 3 4B 38.5%
- 8
LLaVA-v1.5-13B 33%
- 9 llava-v1.5-7b 23.9%
Supplementary surface microstructure analysis
9 models we track have a score here.
- 1
Llama 4 Maverick 90.3%
- 2
InternVL3-38B 74.9%
- 3
InternVL3-8B ↑1 65.9%
- 4
Qwen2.5-VL-32B ↓1 58.4%
- 5
Qwen2.5-VL-7B 49.1%
- 6 llava-v1.5-7b ↑3 41.6%
- 7
Gemma 3 4B 39.3%
- 8
LLaVA-v1.5-13B 35.6%
- 9
Janus-Pro-7B ↓3 33.3%
Supplementary defect type classification
9 models we track have a score here.
- 1
Llama 4 Maverick 66.7%
- 2
InternVL3-38B 63.6%
- 3
Qwen2.5-VL-7B ↑2 60.6%
- 4
Qwen2.5-VL-32B ↓1 60.6%
- 5
Gemma 3 4B ↑2 57.6%
- 6
InternVL3-8B ↓2 57.6%
- 7
Janus-Pro-7B ↓1 48.5%
- 8 llava-v1.5-7b ↑1 33.3%
- 9
LLaVA-v1.5-13B ↓1 27.3%
Supplementary image content analysis
9 models we track have a score here.
- 1
Llama 4 Maverick 62.6%
- 2
InternVL3-38B 41.8%
- 3
InternVL3-8B ↑1 35.9%
- 4
Qwen2.5-VL-32B ↓1 32.5%
- 5
Gemma 3 4B ↑2 25.7%
- 6
Qwen2.5-VL-7B ↓1 24.8%
- 7
Janus-Pro-7B ↓1 21.4%
- 8
LLaVA-v1.5-13B 15.1%
- 9 llava-v1.5-7b 14.6%
Converted VQA
9 models we track have a score here.
- 1
Llama 4 Maverick 75.7%
- 2
InternVL3-38B 55.5%
- 3
InternVL3-8B ↑1 53.2%
- 4
Qwen2.5-VL-32B ↓1 45.3%
- 5
Qwen2.5-VL-7B 39.1%
- 6
Gemma 3 4B ↑1 34.2%
- 7 llava-v1.5-7b ↑2 29.4%
- 8
Janus-Pro-7B ↓2 28.1%
- 9
LLaVA-v1.5-13B ↓1 26.5%
Scores at each effort level
Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.
Source: The MatCha paper
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| 16-shot | 68.1% | ||||
| 2-shot | 66.9% | ||||
| 8-shot | 66.6% | ||||
| 4-shot | 65.9% | ||||
| Zero-shot | 54.8% | ||||
| 2-shot | 52.2% | ||||
| 4-shot | 51.7% | ||||
| Chain of thought | 50.9% | ||||
| 8-shot | 50.4% | ||||
| 16-shot | 50.3% | ||||
| Chain of thought | 50.1% | ||||
| Zero-shot | 49.9% | ||||
| Zero-shot | 48.7% | ||||
| Chain of thought | 45.7% | ||||
| 4-shot | 45.3% | ||||
| 16-shot | 45.1% | ||||
| 8-shot | 44.6% | ||||
| 2-shot | 44.3% | ||||
| 16-shot | 44.1% | ||||
| 8-shot | 44.1% | ||||
| 2-shot | 43.5% | ||||
| Chain of thought | 42.7% | ||||
| 4-shot | 42.1% | ||||
| Chain of thought | 41.6% | ||||
| 4-shot | 41.3% | ||||
| Zero-shot | 41% | ||||
| 2-shot | 40.6% | ||||
| Chain of thought | 40.3% | ||||
| 16-shot | 40.2% | ||||
| 8-shot | 39.9% | ||||
| 16-shot | 38.9% | ||||
| 2-shot | 36.5% | ||||
| Zero-shot | 36.4% | ||||
| 8-shot | 35.1% | ||||
| 4-shot | 34.9% | ||||
| 16-shot | 34.8% | ||||
| Zero-shot | 34.5% | ||||
| Zero-shot | 34.3% | ||||
| Chain of thought | 34.1% | ||||
| 2-shot | 32.5% | ||||
| 4-shot | 32% | ||||
| 2-shot | 29.7% | ||||
| Zero-shot | 27.9% | ||||
| llava-v1.5-7b | 2-shot | 24.1% | |||
| 4-shot | 20.7% | ||||
| llava-v1.5-7b | Zero-shot | 20.5% | |||
| llava-v1.5-7b | 4-shot | 19.9% | |||
| Chain of thought | 15.2% | ||||
| llava-v1.5-7b | Chain of thought | 13.9% | |||
| 8-shot | 0% | ||||
| 16-shot | 0% | ||||
| llava-v1.5-7b | 8-shot | 0% | |||
| llava-v1.5-7b | 16-shot | 0% | |||
| 8-shot | 0% |
Scores at each effort level
Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.
Source: The MatCha paper
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| 16-shot | 68.1% | ||||
| 2-shot | 66.9% | ||||
| 8-shot | 66.6% | ||||
| 4-shot | 65.9% | ||||
| Zero-shot | 54.8% | ||||
| 2-shot | 52.2% | ||||
| 4-shot | 51.7% | ||||
| Chain of thought | 50.9% | ||||
| 8-shot | 50.4% | ||||
| 16-shot | 50.3% | ||||
| Chain of thought | 50.1% | ||||
| Zero-shot | 49.9% | ||||
| Zero-shot | 48.7% | ||||
| Chain of thought | 45.7% | ||||
| 4-shot | 45.3% | ||||
| 16-shot | 45.1% | ||||
| 8-shot | 44.6% | ||||
| 2-shot | 44.3% | ||||
| 16-shot | 44.1% | ||||
| 8-shot | 44.1% | ||||
| 2-shot | 43.5% | ||||
| Chain of thought | 42.7% | ||||
| 4-shot | 42.1% | ||||
| Chain of thought | 41.6% | ||||
| 4-shot | 41.3% | ||||
| Zero-shot | 41% | ||||
| 2-shot | 40.6% | ||||
| Chain of thought | 40.3% | ||||
| 16-shot | 40.2% | ||||
| 8-shot | 39.9% | ||||
| 16-shot | 38.9% | ||||
| 2-shot | 36.5% | ||||
| Zero-shot | 36.4% | ||||
| 8-shot | 35.1% | ||||
| 4-shot | 34.9% | ||||
| 16-shot | 34.8% | ||||
| Zero-shot | 34.5% | ||||
| Zero-shot | 34.3% | ||||
| Chain of thought | 34.1% | ||||
| 2-shot | 32.5% | ||||
| 4-shot | 32% | ||||
| 2-shot | 29.7% | ||||
| Zero-shot | 27.9% | ||||
| llava-v1.5-7b | 2-shot | 24.1% | |||
| 4-shot | 20.7% | ||||
| llava-v1.5-7b | Zero-shot | 20.5% | |||
| llava-v1.5-7b | 4-shot | 19.9% | |||
| Chain of thought | 15.2% | ||||
| llava-v1.5-7b | Chain of thought | 13.9% | |||
| 8-shot | 0% | ||||
| 16-shot | 0% | ||||
| llava-v1.5-7b | 8-shot | 0% | |||
| llava-v1.5-7b | 16-shot | 0% | |||
| 8-shot | 0% |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The MatCha paper
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official MatCha 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 Science and reasoning
- GPQA Diamond In use 198 graduate-level 'Google-proof' multiple-choice questions in biology, chemistry and physics.
Gemini 4 Carbon 97%
- Terminal-Bench-Science In use 70 scientific research-workflow tasks across five science areas, run in Harbor containers. Hidden tests give the resolution rate. GPT-6 Astra 68.1%
- BioMysteryBench In use Bioinformatics problems on anonymized real datasets that require real analysis with tools. Rubric-graded, with human-solvable and human-difficult splits. Claude Fable 5.2 83.7%
- FrontierScience In use Models solve physics, chemistry, and biology problems in an Olympiad track graded by short answer and a Research track graded by 10-point rubrics.
Intern-S2-397B 87%
- HiPhO In use Models solve theory problems from 2024 and 2025 physics olympiad exams, some with diagrams. Official marking schemes grade answers and steps, mapped to medal levels.
Gemini 3 Pro 78.3%
- LABBench2 In use 11 biology research sub-tasks (literature, figures, tables, patents, trials and more) with tool access. Macro-averaged accuracy. GPT-5.2 60.9%
Data audit
9 of 10 checks pass, 1 warning. Look first at the Newest score check. The newest came 19 months ago, in April 2025, for InternVL3-38B. 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 benchmark’s paper (15 entries).
- Pass Models scored 9 models we track, plus 6 other entries on its leaderboard.
- Warning Newest score The newest came 19 months ago, in April 2025, for InternVL3-38B.
- Note Frontier models None of the top 10 has a score. The strongest tested is Llama 4 Maverick (#153).
- Pass Headroom The best is 68.1%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on arxiv.org: 54 rows for models we track and 6 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.