Massive Multi-discipline Multimodal Understanding (MMMU)
A broad multimodal reasoning benchmark spanning charts, diagrams, tables, and academic visual question answering.
Top models on MMMU — November 21, 2026
As of November 21, 2026, Gemini 3 Pro leads the MMMU leaderboard with 87.2% , followed by GPT-5.2 (86.7%) and Qwen3.6 Plus (86.0%).
Gemini 3 Pro
GPT-5.2
OpenAI
Another lab’s comparison tableQwen3.6 Plus
Alibaba
Carried from an upstream page32 modelsMultimodal & Grounded2% of Multimodal reference weightRefreshingUpdated November 21, 2026
Leaderboard (32 models)
None of the 32 results on this page has a documented run by the benchmark or an independent evaluator: 24 reported by the labs that built the models, 4 reported in another publisher’s document and 4 carried from an upstream page, where who ran them is not recorded.
| Rank | Model / configuration | Score | Who ran it | Parameters (B) | Open / closed |
|---|---|---|---|---|---|
| 1 | Gemini 3 ProGoogleAnother lab’s comparison table | 87.2% | Another lab’s comparison table | Not reported | Closed |
| 2 | GPT-5.2OpenAIAnother lab’s comparison table | 86.7% | Another lab’s comparison table | Not reported | Closed |
| 3 | Qwen3.6 PlusAlibabaCarried from an upstream page | 86.0% | Carried from an upstream page | Not reported | Closed |
| 4 | Qwen3.5 397BAlibabaReported by Alibaba | 85.0% | Reported by Alibaba | Not reported | Open |
| 5 | Kimi K2.5Moonshot AIAnother lab’s comparison table | 84.3% | Another lab’s comparison table | Not reported | Open |
| 6 | Qwen3.5-122B-A10BAlibabaCarried from an upstream page | 83.9% | Carried from an upstream page | Not reported | Open |
| 7 | Qwen3.6-27BAlibabaReported by Alibaba | 82.9% | Reported by Alibaba | Not reported | Open |
| 8 | Qwen3.5-27BAlibabaCarried from an upstream page | 82.3% | Carried from an upstream page | Not reported | Open |
| 9 | Qwen3.6-35B-A3BAlibabaReported by Alibaba | 81.7% | Reported by Alibaba | Not reported | Open |
| 10 | Qwen3.5-35B-A3BAlibabaCarried from an upstream page | 81.4% | Carried from an upstream page | Not reported | Open |
| 11 | Claude Opus 4.5AnthropicAnother lab’s comparison table | 80.7% | Another lab’s comparison table | Not reported | Closed |
| 12 | Qwen3-VL-235B-A22B InstructAlibabaReported by Alibaba | 78.7% | Reported by Alibaba | Not reported | Open |
| 13 | Claude 4 Opus ThinkingAnthropicReported by Anthropic | 76.5% | Reported by Anthropic | Not reported | Closed |
| 14 | GLM-4.6VZ.AIReported by Z.AI | 76.0% | Reported by Z.AI | Not reported | Open |
| 15 | Command A+CohereReported by Cohere | 75.1% | Reported by Cohere | Not reported | Open |
| 16 | Claude 3.7 Sonnet ThinkingAnthropicReported by Anthropic | 75.0% | Reported by Anthropic | Not reported | Closed |
| 17 | Claude 4 OpusAnthropicReported by Anthropic | 73.7% | Reported by Anthropic | Not reported | Closed |
| 18 | Gemini 2.5 Flash-LiteGoogleReported by Google | 72.9% | Reported by Google | Not reported | Closed |
| 19 | Claude 3.7 SonnetAnthropicReported by Anthropic | 71.8% | Reported by Anthropic | Not reported | Closed |
| 20 | GLM-4.6V-FlashZ.AIReported by Z.AI | 71.1% | Reported by Z.AI | Not reported | Open |
| 21 | Nemotron 3 Nano Omni 30B A3BNVIDIAReported by NVIDIA | 70.8% | Reported by NVIDIA | Not reported | Open |
| 22 | Gemini 2.0 FlashGoogleReported by Google | 69.3% | Reported by Google | Not reported | Closed |
| 23 | MiniMax-VL-01MiniMaxReported by MiniMax | 68.5% | Reported by MiniMax | Not reported | Open |
| 24 | Grok 2xAIReported by xAI | 66.1% | Reported by xAI | Not reported | Closed |
| 25 | Gemini 2.0 Flash-LiteGoogleReported by Google | 65.1% | Reported by Google | Not reported | Closed |
| 26 | Mistral Small 3.1MistralReported by Mistral | 64.0% | Reported by Mistral | Not reported | Open |
| 27 | Grok 2 minixAIReported by xAI | 63.2% | Reported by xAI | Not reported | Closed |
| 28 | Gemini 1.5 FlashGoogleReported by Google | 58.3% | Reported by Google | Not reported | Closed |
| 29 | LFM3-VL-2BLiquidAIReported by LiquidAI | 55.7% | Reported by LiquidAI | Not reported | Open |
| 30 | LFM2.5-VL-3BLiquidAIReported by LiquidAI | 48.4% | Reported by LiquidAI | Not reported | Open |
| 31 | ZAYA1-VL-8BZyphraReported by Zyphra | 46.0% | Reported by Zyphra | Not reported | Open |
| 32 | LFM2.5-VL-450MLiquidAIReported by LiquidAI | 32.7% | Reported by LiquidAI | Not reported | Open |
- Lab-reported
- Benchmark-run
- Independent
MMMU
BenchLM mirrors this Vals AI leaderboard for reference only. It is excluded from BenchLM overall and category rankings.
| Source rank | Source model or configuration | MMMU score | Source setup | Parameters (B) | Open / closed |
|---|---|---|---|---|---|
| 1 | Claude Fable 5.1Anthropic | 90.64% | Vals AIOriginal result | Not reported | Closed |
| 2 | Claude Opus 5Anthropic | 89.88% | Vals AIOriginal result | Not reported | Closed |
| 3 | Claude Fable 5Anthropic | 89.31% | Vals AIOriginal result | Not reported | Closed |
| 4 | Gemini 3.8 FlashGoogle | 89.08% | high reasoningOriginal result | Not reported | Closed |
| 5 | Gemini 3.7 FlashGoogle | 88.96% | high reasoningOriginal result | Not reported | Closed |
| 6 | GPT-5.6 SolOpenAI | 88.84% | max reasoningOriginal result | Not reported | Closed |
| 7 | Gemini 3.6 FlashGoogle | 88.38% | high reasoningOriginal result | Not reported | Closed |
| 8 | Gemini 3.5 FlashGoogle | 88.27% | high reasoningOriginal result | Not reported | Closed |
| 9 | GPT-5.5OpenAI | 88.27% | xhigh reasoningOriginal result | Not reported | Closed |
| 10 | Gemini 3.1 Pro PreviewGoogle | 88.21% | high reasoningOriginal result | Not reported | Unknown |
| 11 | Kimi K3Moonshot AI | 88.15% | Vals AIOriginal result | Not reported | Open |
| 12 | Qwen3.8 MaxAlibaba | 88.03% | Vals AIOriginal result | Not reported | Open |
| 13 | Gemini 3 Flash PreviewGoogle | 87.63% | high reasoningOriginal result | Not reported | Unknown |
| 14 | Gemini 3 Pro PreviewGoogle | 87.51% | high reasoningOriginal result | Not reported | Unknown |
| 15 | GPT-5.4OpenAI | 87.51% | xhigh reasoningOriginal result | Not reported | Closed |
| 16 | Muse SparkMeta | 87.40% | Vals AIOriginal result | Not reported | Closed |
| 17 | GPT-5.2OpenAI | 86.67% | xhigh reasoningOriginal result | Not reported | Closed |
| 18 | Muse Spark 1.1Meta | 86.59% | xhigh reasoningOriginal result | Not reported | Closed |
| 19 | Claude Opus 4.8Anthropic | 86.59% | Vals AIOriginal result | Not reported | Closed |
| 20 | GPT-5.6 TerraOpenAI | 86.47% | max reasoningOriginal result | Not reported | Closed |
| 21 | Kimi K2.6Moonshot AI | 86.30% | Vals AIOriginal result | Not reported | Open |
| 22 | Muse Spark 1.2Meta | 86.13% | xhigh reasoningOriginal result | Not reported | Closed |
| 23 | GLM-5.3-FlashZ.AI | 86.01% | max reasoningOriginal result | Not reported | Open |
| 24 | Claude Opus 4.7Anthropic | 85.55% | Vals AIOriginal result | Not reported | Closed |
| 25 | GPT-5.6 LunaOpenAI | 85.03% | max reasoningOriginal result | Not reported | Closed |
| 26 | Kimi K2.5 ThinkingMoonshot AI | 84.33% | Vals AIOriginal result | Not reported | Unknown |
| 27 | Qwen3.6 PlusAlibaba | 84.16% | Vals AIOriginal result | Not reported | Closed |
| 28 | Qwen3.8-27BAlibaba | 83.93% | xhigh reasoningOriginal result | Not reported | Open |
| 29 | Claude Opus 4.6 (Adaptive)Anthropic | 83.87% | Vals AIOriginal result | Not reported | Closed |
| 30 | Gemini 3.5 Flash-LiteGoogle | 83.58% | high reasoningOriginal result | Not reported | Closed |
| 31 | Claude Sonnet 4.6Anthropic | 83.58% | Vals AIOriginal result | Not reported | Closed |
| 32 | Grok 4.20 0309 ReasoningxAI | 83.47% | Vals AIOriginal result | Not reported | Unknown |
| 33 | GPT-5.1OpenAI | 83.18% | high reasoningOriginal result | Not reported | Closed |
| 34 | Grok 4.3xAI | 83.06% | Vals AIOriginal result | Not reported | Closed |
| 35 | Claude Sonnet 5Anthropic | 83.01% | Vals AIOriginal result | Not reported | Closed |
| 36 | Claude Opus 4.5 ThinkingAnthropic | 82.95% | Vals AIOriginal result | Not reported | Closed |
| 37 | Gemini 3.1 Flash Lite PreviewGoogle | 82.49% | high reasoningOriginal result | Not reported | Unknown |
| 38 | Qwen3.5 FlashAlibaba | 81.91% | Vals AIOriginal result | Not reported | Closed |
| 39 | GPT-5OpenAI | 81.50% | high reasoningOriginal result | Not reported | Closed |
| 40 | Gemini 2.5 Pro Exp 03 25Google | 81.34% | Vals AIOriginal result | Not reported | Unknown |
| 41 | MiniMax M3MiniMax | 81.16% | Vals AIOriginal result | Not reported | Open |
| 42 | Claude Opus 4.5Anthropic | 81.10% | Vals AIOriginal result | Not reported | Closed |
| 43 | Gemini 2.5 Flash Preview 09 2025 ThinkingGoogle | 80.75% | Vals AIOriginal result | Not reported | Unknown |
| 44 | o3OpenAI | 80.42% | high reasoningOriginal result | Not reported | Closed |
| 45 | MiMo-V2.5Xiaomi | 80.00% | Vals AIOriginal result | Not reported | Closed |
| 46 | o4-miniOpenAI | 79.67% | high reasoningOriginal result | Not reported | Closed |
| 47 | Gemini 2.5 Flash Preview 09 2025Google | 79.48% | Vals AIOriginal result | Not reported | Unknown |
| 48 | Claude Sonnet 4.5 ThinkingAnthropic | 79.31% | Vals AIOriginal result | Not reported | Closed |
| 49 | GPT-5.4 miniOpenAI | 79.25% | xhigh reasoningOriginal result | Not reported | Closed |
| 50 | GPT-5 miniOpenAI | 78.91% | high reasoningOriginal result | Not reported | Closed |
| 51 | InklingThinking Machines Lab | 78.50% | 0.99 reasoningOriginal result | Not reported | Open |
| 52 | Claude Opus 4.1 20250805 ThinkingAnthropic | 77.51% | Vals AIOriginal result | Not reported | Unknown |
| 53 | o1OpenAI | 77.41% | high reasoningOriginal result | Not reported | Closed |
| 54 | Grok 4 0709xAI | 76.27% | Vals AIOriginal result | Not reported | Unknown |
| 55 | Gemini 2.5 Flash Lite Preview 09 2025 ThinkingGoogle | 75.43% | Vals AIOriginal result | Not reported | Unknown |
| 56 | Claude 3.7 Sonnet ThinkingAnthropic | 75.10% | Vals AIOriginal result | Not reported | Closed |
| 57 | Claude Sonnet 4 20250514 ThinkingAnthropic | 74.93% | Vals AIOriginal result | Not reported | Unknown |
| 58 | Claude Opus 4.1Anthropic | 73.72% | Vals AIOriginal result | Not reported | Unknown |
| 59 | GPT-5.4 nanoOpenAI | 73.58% | high reasoningOriginal result | Not reported | Closed |
| 60 | Claude Opus 4Anthropic | 73.31% | Vals AIOriginal result | Not reported | Unknown |
| 61 | Grok 4 Fast (Reasoning)xAI | 72.78% | Vals AIOriginal result | Not reported | Closed |
| 62 | Grok 4.1 Fast (Reasoning)xAI | 72.66% | Vals AIOriginal result | Not reported | Closed |
| 63 | Gemini 2.5 Flash Lite Preview 09 2025Google | 72.54% | Vals AIOriginal result | Not reported | Unknown |
| 64 | Claude Sonnet 4Anthropic | 72.39% | Vals AIOriginal result | Not reported | Unknown |
| 65 | GPT-4.1OpenAI | 72.39% | high reasoningOriginal result | Not reported | Closed |
| 66 | Gemini 2.5 Flash Preview 04 17 ThinkingGoogle | 71.92% | Vals AIOriginal result | Not reported | Unknown |
| 67 | Llama4 Maverick Instruct BasicFireworks AI | 71.69% | Vals AIOriginal result | Not reported | Unknown |
| 68 | Claude 3.7 SonnetAnthropic | 71.52% | Vals AIOriginal result | Not reported | Closed |
| 69 | GPT-5 nanoOpenAI | 70.94% | high reasoningOriginal result | Not reported | Closed |
| 70 | GPT-4.1 miniOpenAI | 70.54% | high reasoningOriginal result | Not reported | Closed |
| 71 | Gemini 2.0 Flash 001Google | 69.79% | Vals AIOriginal result | Not reported | Unknown |
| 72 | Claude 3.5 SonnetAnthropic | 68.80% | Vals AIOriginal result | Not reported | Closed |
| 73 | Gemini 2.5 Flash Preview 04 17Google | 68.00% | Vals AIOriginal result | Not reported | Unknown |
| 74 | Mistral Large 2512Mistral AI | 66.19% | Vals AIOriginal result | Not reported | Unknown |
| 75 | Gemini 1.5 Pro 002Google | 65.51% | Vals AIOriginal result | Not reported | Unknown |
| 76 | Magistral Small 2509Mistral AI | 65.20% | Vals AIOriginal result | Not reported | Unknown |
| 77 | Magistral Medium 2509Mistral AI | 64.57% | Vals AIOriginal result | Not reported | Unknown |
| 78 | GPT-4oOpenAI | 64.01% | high reasoningOriginal result | Not reported | Closed |
| 79 | Grok 4.1 Fast Non ReasoningxAI | 63.70% | Vals AIOriginal result | Not reported | Unknown |
| 80 | Grok 4 Fast Non ReasoningxAI | 63.41% | Vals AIOriginal result | Not reported | Unknown |
| 81 | Mistral Medium 2505Mistral AI | 62.97% | Vals AIOriginal result | Not reported | Unknown |
| 82 | GPT-4oOpenAI | 62.16% | high reasoningOriginal result | Not reported | Closed |
| 83 | Grok 4.5xAI | 61.79% | high reasoningOriginal result | Not reported | Closed |
| 84 | Mistral Small 2503Mistral AI | 60.08% | Vals AIOriginal result | Not reported | Unknown |
| 85 | Meta Llama Llama 4 Scout 17B 16E InstructTogether AI | 58.75% | Vals AIOriginal result | Not reported | Unknown |
| 86 | Grok 2 Vision 1212xAI | 57.25% | Vals AIOriginal result | Not reported | Unknown |
| 87 | Gemini 1.5 Flash 002Google | 57.19% | Vals AIOriginal result | Not reported | Unknown |
| 88 | GPT-4o miniOpenAI | 56.56% | high reasoningOriginal result | Not reported | Closed |
| 89 | GPT-4.1 nanoOpenAI | 55.05% | high reasoningOriginal result | Not reported | Closed |
| 90 | Meta Llama Llama 3.2 90B Vision Instruct TurboTogether AI | 48.06% | Vals AIOriginal result | Not reported | Unknown |
| 91 | Claude Haiku 4.5 ThinkingAnthropic | 46.07% | Vals AIOriginal result | Not reported | Closed |
| 92 | Meta Llama Llama 3.2 11B Vision Instruct TurboTogether AI | 38.82% | Vals AIOriginal result | Not reported | Unknown |
| 93 | Qwen3.5 Plus ThinkingAlibaba | 22.77% | Vals AIOriginal result | Not reported | Unknown |
According to BenchLM.ai, Gemini 3 Pro leads the MMMU benchmark with a score of 87.2%, followed by GPT-5.2 (86.7%) and Qwen3.6 Plus (86.0%). The top models are clustered within 1.2 points, suggesting this benchmark is nearing saturation for frontier models.
32 models have been evaluated on MMMU. The benchmark falls in the Multimodal & Grounded category. BenchAlign v6.0 gives MMMU 2% of the Multimodal reference weight, so it moves the Multimodal leaderboard. Reference weights are relative weights in the calibrated model, not fixed shares of a score.
About MMMU
Year
2024
Tasks
Multimodal academic reasoning
Format
Image + text question answering
Difficulty
Frontier multimodal
MMMU is the base benchmark family behind later MMMU-Pro variants. It measures whether a model can answer expert-style questions that require combining visual understanding with domain knowledge and reasoning.
Freshness and provenance
Version
MMMU 2024
Refresh cadence
Annual
Staleness state
Refreshing
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does MMMU measure?
A broad multimodal reasoning benchmark spanning charts, diagrams, tables, and academic visual question answering.
Which model scores highest on MMMU?
Gemini 3 Pro by Google currently leads with a score of 87.2% on MMMU.
How many models are evaluated on MMMU?
32 AI models have published results on MMMU in the BenchLM catalog.
Compare top models on MMMU
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 7,000+ subscribers.
One email each week. Unsubscribe anytime.