MMMU-Pro
Harder, filtered variant of MMMU for multimodal college-level reasoning, with more answer options and questions embedded in images.
Actively evaluated by Artificial Analysis on current frontier models; no saturation evidence recorded.
Performance Timeline
Longitudinal progression of model scores against human baselines.Performance & Historical Trajectory
Empirical score progression across model release dates and evaluation rounds.
| Model | Score | Date | Source Type | Provenance |
|---|---|---|---|---|
| Magistral Small 1.2 | 55.5% | 2026-11-15 | independent | Source ↗ |
| Llama 4 Maverick | 62.1% | 2026-11-15 | independent | Source ↗ |
| Nemotron 3 Nano Omni 30B A3B Reasoning | 53.2% | 2026-11-15 | independent | Source ↗ |
| Magistral Medium 1.2 | 59.7% | 2026-11-15 | independent | Source ↗ |
| DeepSeek V4 Flash Vision | 74.8% | 2026-11-15 | independent | Source ↗ |
| Gemma 4 E2B | 41.8% | 2026-11-15 | independent | Source ↗ |
| Gemini 3.8 Flash | 85.6% | 2026-11-15 | independent | Source ↗ |
| Inkling Small | 74% | 2026-11-15 | independent | Source ↗ |
| Ministral 3 14B | 49.8% | 2026-11-15 | independent | Source ↗ |
| Ministral 3 8B | 46% | 2026-11-15 | independent | Source ↗ |
| Qwen3.5 Omni Flash | 64.7% | 2026-11-15 | independent | Source ↗ |
| NVIDIA Nemotron Nano 12B v2 VL | 52.9% | 2026-11-15 | independent | Source ↗ |
| Gemma 4 26B A4B | 69.2% | 2026-11-15 | independent | Source ↗ |
| Ministral 3 3B | 38.1% | 2026-11-15 | independent | Source ↗ |
| Qwen3.6 35B A3B | 71% | 2026-11-15 | independent | Source ↗ |
| Kimi K3.1 Pro | 83.9% | 2026-11-15 | independent | Source ↗ |
| Nova 2.0 Lite | 63.8% | 2026-11-15 | independent | Source ↗ |
| Llama 3.2 Instruct 11B (Vision) | 29.3% | 2026-11-15 | independent | Source ↗ |
| Qwen3 Omni 30B A3B (Reasoning) | 60.2% | 2026-11-15 | independent | Source ↗ |
| Agnes 2.5 Pro Beta | 75.4% | 2026-11-15 | independent | Source ↗ |
| Inkling 2 Small | 81.6% | 2026-11-15 | independent | Source ↗ |
| Mistral Medium 3.5 | 64.9% | 2026-11-15 | independent | Source ↗ |
| Agnes 2.5 Pro Theta | 78.4% | 2026-11-15 | independent | Source ↗ |
| MiniMax M3.1 Flash | 79.1% | 2026-11-15 | independent | Source ↗ |
| Gemma 4 31B | 73.4% | 2026-11-15 | independent | Source ↗ |
| DiffusionGemma 26B A4B | 66.5% | 2026-11-15 | independent | Source ↗ |
| Gemini 3.1 Pro | 82.4% | 2026-11-15 | independent | Source ↗ |
| GPT-5.3-Codex | 78.5% | 2026-11-15 | independent | Source ↗ |
| Qwen3.8 Max (0902) | 82.8% | 2026-11-15 | independent | Source ↗ |
| Molmo 7B-D | 24.5% | 2026-11-15 | independent | Source ↗ |
| Llama 4 Scout | 52.9% | 2026-11-15 | independent | Source ↗ |
| Step3 VL 10B | 64% | 2026-11-15 | independent | Source ↗ |
| Qwen3.7 Plus | 80.5% | 2026-11-15 | independent | Source ↗ |
| ERNIE 5.0 Thinking | 64.6% | 2026-11-15 | independent | Source ↗ |
| Llama 3.2 Instruct 90B (Vision) | 39.5% | 2026-11-15 | independent | Source ↗ |
| Qwen3.8-Flash-Next | 78.6% | 2026-11-15 | independent | Source ↗ |
| Qwen3 Omni 30B A3B Instruct | 55.5% | 2026-11-15 | independent | Source ↗ |
| GPT-6 Luna | 79.7% | 2026-11-15 | independent | Source ↗ |
| o3 | 70.1% | 2026-11-15 | independent | Source ↗ |
| Mistral Small 4 | 56.8% | 2026-11-15 | independent | Source ↗ |
| MiMo-V2.6-Flash | 73.1% | 2026-11-15 | independent | Source ↗ |
| Qwen3.5 397B A17B | 77.3% | 2026-11-15 | independent | Source ↗ |
| Inkling | 73.5% | 2026-11-15 | independent | Source ↗ |
| Nova 2.0 Pro | 64.5% | 2026-11-15 | independent | Source ↗ |
| GPT-6.1 Terra | 84.7% | 2026-11-15 | independent | Source ↗ |
| GPT-6 Astra | 86.9% | 2026-11-15 | independent | Source ↗ |
| Qwen3.5 2B | 42.7% | 2026-11-15 | independent | Source ↗ |
| Step 5 | 76.4% | 2026-11-15 | independent | Source ↗ |
| Gemini 3.5 Flash-Lite | 79% | 2026-11-15 | independent | Source ↗ |
| GPT-6.1 Sol | 86% | 2026-11-15 | independent | Source ↗ |
| Apodex 1.1 | 79.2% | 2026-11-15 | independent | Source ↗ |
| Gemma 4 12B | 62% | 2026-11-15 | independent | Source ↗ |
| Qwen3.5-122B-A10B | 70.3% | 2026-11-15 | independent | Source ↗ |
| Nova 2.0 Omni | 61.9% | 2026-11-15 | independent | Source ↗ |
| LFM2.5-VL-1.6B | 26.5% | 2026-11-15 | independent | Source ↗ |
| Muse Glimmer | 74.3% | 2026-11-15 | independent | Source ↗ |
| DeepSeek V4.1 Flash 1017 | 80.6% | 2026-11-15 | independent | Source ↗ |
| Gemma 4 E4B | 51.2% | 2026-11-15 | independent | Source ↗ |
| Doubao Seed Code | 68.1% | 2026-11-15 | independent | Source ↗ |
| MiniCPM-V 4.6 1.3B | 37.9% | 2026-11-15 | independent | Source ↗ |
| Ling-3.0-flash-VL | 79% | 2026-11-15 | independent | Source ↗ |
| Qwen3.5 0.8B | 25.7% | 2026-11-15 | independent | Source ↗ |
| Molmo2-8B | 37.5% | 2026-11-15 | independent | Source ↗ |
| Qwen3.5 Omni Plus | 70.5% | 2026-11-15 | independent | Source ↗ |
| Muse Spark Open | 80.9% | 2026-11-15 | independent | Source ↗ |
| Qwen3.8 27B | 76.3% | 2026-11-15 | independent | Source ↗ |
| Qwen3.5 4B | 62.1% | 2026-11-15 | independent | Source ↗ |
| Claude Opus 5.5 | 87.7% | 2026-11-15 | independent | Source ↗ |
| Command A+ | 63.2% | 2026-11-15 | independent | Source ↗ |
| Phi-4-multimodal | 14.5% | 2026-11-15 | independent | Source ↗ |
| MiniMax M3 | 78.6% | 2026-11-15 | independent | Source ↗ |
| Qwen3.5-9B | 66.8% | 2026-11-15 | independent | Source ↗ |
Human Baseline & Difficulty Horizon
Calibrated human reference points, specialist benchmarks, and ceiling thresholds.No peer-reviewed human baseline has been published for this benchmark.
Metric & Scoring Methodology
Verification protocols, aggregation formulas, and specialized metric variants.accuracy (%)Dataset & Compute Cost
Evaluation volume, public availability, API pricing, and local hardware requirements.$5 – $20 USD for full benchmark evaluation run on frontier APIs.
1x NVIDIA RTX 4090 (24GB) or A100 (40GB/80GB) via vLLM / SGLang
How to Run & Reproduce
Standardized evaluation protocols, CLI commands, and reproducible runner templates.lm_eval --model hf --model_args pretrained=<model_path> --tasks mmmu-pro --batch_size autoopencompass --datasets mmmu-pro --models <model_config># Standard API Evaluation Loop
from openai import OpenAI
client = OpenAI()
response = client.chat.completions.create(
model="gpt-4o",
messages=[{"role": "user", "content": prompt}],
temperature=0.0,
)When publishing results for MMMU-Pro, always report the exact prompt template, few-shot exemplar ordering, sampling temperature (temperature=0), maximum reasoning budget tokens, and the precise timestamped model snapshot ID.
Contamination & Memorization Analysis
Audit of pretraining exposure risks, memorization vectors, and refresh policies.Static fixed snapshot
Public on web / HuggingFace