Benchmark Registry

MMMU-Pro

MMMU-Pro results reported for 64 models across 5 versions. Updated 2026-10-07.

Report an issue

Versions

5 versions · Metric: MMMU-Pro accuracy

VersionReleasedModelsResults
MMMU-Pro With toolsConfiguration: With toolsMarch 5, 202666
MMMU-Pro With PythonConfiguration: With PythonDecember 11, 202533
MMMU-Pro 1,730 questions — Mercor single-shotConfiguration: Mercor single-shot agentSeptember 4, 20243234
MMMU-Pro No toolsConfiguration: No toolsSeptember 4, 20241718
MMMU-Pro Vals AI harnessConfiguration: Vals AI harnessSeptember 4, 20245560

Settings

Each score is labelled with the setting its source states.

Overall (Standard 10 options + Vision)
The average of the Standard setting with augmented options (usually 10) and the Vision-only input setting. The paper defines this as the overall MMMU-Pro score (Section 3.1; Table 1). Source
Standard (10 options)
The Standard setting with augmented options (usually 10) only. Source
Vision
The Vision-only input setting only: each question and its options are embedded in a screenshot or photo. Source
Standard (4 options)
The Standard setting without augmented options (usually 4). The paper reports it only for comparison; it is not part of the overall score. Source
Not stated
The source reports an MMMU-Pro score without saying which setting it covers. Source

Recently reported results: MMMU-Pro With tools

Most recent configuration: MMMU-Pro With tools

ModelScoreSource
Kimi K3 (max)83.4%huggingface.coMMMU-Pro accuracySetting: Not statedReported Model cardMoonshot AISelf-reportedRecord
GPT-5.6 Terra82.0%openai.comMMMU-Pro accuracySetting: Not statedReported Launch postOpenAISelf-reportedRecord
GPT-5.6 Luna79.5%openai.comMMMU-Pro accuracySetting: Not statedReported Launch postOpenAISelf-reportedRecord
GPT-5.6 Sol84.6%openai.comMMMU-Pro accuracySetting: Not statedReported Launch postOpenAISelf-reportedRecord
GPT-5.5 (xhigh)83.2%openai.comMMMU-Pro accuracySetting: Not statedReported Launch postOpenAISelf-reportedRecord