MMBench

Indexed results from MMBench. Each ranking retains its own metric and evaluation conditions.

image_scene

mmBench · MMBench · image_scene · percent · Higher is better

  1. vikhyatk/moondream2vikhyatk97.7%
  2. Qwen: Qwen2.5 VL 72B Instructqwen94.6%
  3. DeepSeek: DeepSeek-VL2deepseek93.8%
  4. HuggingFaceM4/Idefics3-8B-Llama3huggingfacem493.8%
  5. Qwen/Qwen2-VL-72B-Instructqwen93.8%
  6. moonshotai/Kimi-VL-A3B-Instructmoonshotai93%
  7. OpenGVLab/InternVL3-2Bopengvlab93%
  8. ibm-granite/granite-vision-3.2-2bibm-granite92.2%
  9. microsoft/Phi-3.5-vision-instructmicrosoft92.2%
  10. Qwen/Qwen2.5-VL-3B-Instructqwen91.5%
  11. Qwen/Qwen2.5-VL-7B-Instructqwen91.5%
  12. tiiuae/falcon-11B-vlmtiiuae91.5%

LR

mmBench · MMBench · LR · percent · Higher is better

  1. Qwen: Qwen2.5 VL 72B Instructqwen83.2%
  2. Qwen/Qwen2.5-VL-32B-Instructqwen82.1%
  3. YannQi/R-4Byannqi80.4%
  4. Qwen/Qwen2-VL-72B-Instructqwen79.9%
  5. XiaomiMiMo/MiMo-VL-7B-RLxiaomi78.8%
  6. moonshotai/Kimi-VL-A3B-Thinking-2506moonshotai76.6%
  7. OpenGVLab/InternVL2-40Bopengvlab76.6%
  8. NCSOFT/VARCO-VISION-2.0-14Bncsoft76.1%
  9. Google: Gemma 3 27Bgoogle72.3%
  10. Qwen/Qwen2.5-VL-7B-Instructqwen71.2%
  11. OpenGVLab/InternVL2_5-8Bopengvlab70.1%
  12. DeepSeek: DeepSeek-VL2deepseek69.6%

ocr

mmBench · MMBench · ocr · percent · Higher is better

  1. Qwen: Qwen2.5 VL 72B Instructqwen100%
  2. NCSOFT/VARCO-VISION-2.0-14Bncsoft98.9%
  3. YannQi/R-4Byannqi98.9%
  4. Qwen/Qwen2-VL-7B-Instructqwen97.8%
  5. nvidia/NVLM-D-72Bnvidia96.7%
  6. OpenGVLab/InternVL2-40Bopengvlab96.7%
  7. OpenGVLab/InternVL2_5-8Bopengvlab96.7%
  8. OpenGVLab/InternVL3-8Bopengvlab96.7%
  9. Qwen/Qwen2.5-VL-32B-Instructqwen96.7%
  10. Qwen/Qwen2-VL-72B-Instructqwen96.7%
  11. Google: Gemma 3 27Bgoogle95.6%
  12. Qwen/Qwen2.5-VL-3B-Instructqwen95.6%

image_style

mmBench · MMBench · image_style · percent · Higher is better

  1. DeepSeek: DeepSeek-VL2deepseek85.9%
  2. OpenGVLab/InternVL2_5-4Bopengvlab84.8%
  3. XiaomiMiMo/MiMo-VL-7B-RLxiaomi84.8%
  4. Google: Gemma 3 27Bgoogle83.7%
  5. OpenGVLab/InternVL2-40Bopengvlab83.7%
  6. Qwen: Qwen2.5 VL 72B Instructqwen83.7%
  7. OpenGVLab/InternVL2_5-8Bopengvlab82.6%
  8. OpenGVLab/InternVL3-2Bopengvlab82.6%
  9. mistralai/Pixtral-12B-2409mistralai81.5%
  10. NCSOFT/VARCO-VISION-2.0-14Bncsoft81.5%
  11. openbmb/MiniCPM-V-4openbmb81.5%
  12. Qwen/Qwen2.5-VL-7B-Instructqwen81.5%

historical_figure

mmBench · MMBench · historical_figure · percent · Higher is better

  1. Qwen/Qwen2-VL-72B-Instructqwen60%
  2. Qwen: Qwen2.5 VL 72B Instructqwen57.1%
  3. nvidia/NVLM-D-72Bnvidia51.4%
  4. moonshotai/Kimi-VL-A3B-Instructmoonshotai48.6%
  5. rhymes-ai/Ariarhymes-ai37.1%
  6. DeepSeek: DeepSeek-VL2deepseek31.4%
  7. Qwen/Qwen2.5-VL-3B-Instructqwen28.6%
  8. Qwen/Qwen2-VL-7B-Instructqwen28.6%
  9. 5CD-AI/Vintern-3B-beta5cd-ai25.7%
  10. microsoft/Phi-3.5-vision-instructmicrosoft25.7%
  11. Qwen/Qwen2.5-VL-7B-Instructqwen25.7%
  12. Google: Gemma 3 12Bgoogle20%

image_emotion

mmBench · MMBench · image_emotion · percent · Higher is better

  1. ibm-granite/granite-vision-3.2-2bibm-granite79.5%
  2. Qwen/Qwen2-VL-72B-Instructqwen79.5%
  3. OpenGVLab/InternVL2-40Bopengvlab78.3%
  4. Qwen: Qwen2.5 VL 72B Instructqwen78.3%
  5. HuggingFaceM4/Idefics3-8B-Llama3huggingfacem477.1%
  6. OpenGVLab/InternVL3-8Bopengvlab75.9%
  7. Qwen/Qwen2.5-VL-7B-Instructqwen75.9%
  8. tiiuae/falcon-11B-vlmtiiuae75.9%
  9. DeepSeek: DeepSeek-VL2deepseek74.7%
  10. Google: Gemma 3 4Bgoogle74.7%
  11. microsoft/Phi-3.5-vision-instructmicrosoft74.7%
  12. vikhyatk/moondream2vikhyatk73.5%

social_relation

mmBench · MMBench · social_relation · percent · Higher is better

  1. OpenGVLab/InternVL3-2Bopengvlab91.2%
  2. Qwen: Qwen2.5 VL 72B Instructqwen91.2%
  3. Qwen/Qwen2.5-VL-3B-Instructqwen90.1%
  4. Qwen/Qwen2.5-VL-7B-Instructqwen90.1%
  5. Google: Gemma 3 12Bgoogle89%
  6. OpenGVLab/InternVL2_5-4Bopengvlab89%
  7. OpenGVLab/InternVL3-8Bopengvlab89%
  8. Qwen/Qwen2-VL-72B-Instructqwen89%
  9. Qwen/Qwen2-VL-7B-Instructqwen89%
  10. Google: Gemma 3 4Bgoogle87.9%
  11. openbmb/MiniCPM-V-4openbmb87.9%
  12. Qwen/Qwen2.5-VL-32B-Instructqwen86.8%

action_recognition

mmBench · MMBench · action_recognition · percent · Higher is better

  1. ibm-granite/granite-vision-3.2-2bibm-granite95.5%
  2. OpenGVLab/InternVL3-8Bopengvlab95.5%
  3. Qwen/Qwen2-VL-72B-Instructqwen95.5%
  4. Google: Gemma 3 4Bgoogle94.3%
  5. nvidia/NVLM-D-72Bnvidia94.3%
  6. OpenGVLab/InternVL2-40Bopengvlab94.3%
  7. OpenGVLab/InternVL3-2Bopengvlab94.3%
  8. 5CD-AI/Vintern-3B-beta5cd-ai93.2%
  9. DeepSeek: DeepSeek-VL2deepseek93.2%
  10. Google: Gemma 3 12Bgoogle93.2%
  11. HuggingFaceTB/SmolVLM-Instructhuggingfacetb93.2%
  12. tiiuae/falcon-11B-vlmtiiuae93.2%

image_quality

mmBench · MMBench · image_quality · percent · Higher is better

  1. Qwen: Qwen2.5 VL 72B Instructqwen90.6%
  2. Qwen/Qwen2.5-VL-7B-Instructqwen87.1%
  3. OpenGVLab/InternVL2-40Bopengvlab85.9%
  4. DeepSeek: DeepSeek-VL2deepseek84.7%
  5. Qwen/Qwen2-VL-72B-Instructqwen84.7%
  6. rhymes-ai/Ariarhymes-ai84.7%
  7. moonshotai/Kimi-VL-A3B-Instructmoonshotai82.4%
  8. nvidia/NVLM-D-72Bnvidia82.4%
  9. OpenGVLab/InternVL3-8Bopengvlab82.4%
  10. Qwen/Qwen2-VL-7B-Instructqwen82.4%
  11. OpenGVLab/InternVL3-2Bopengvlab80%
  12. Qwen/Qwen2.5-VL-3B-Instructqwen74.1%

Overall

mmBench · MMBench · Overall · percent · Higher is better

  1. Qwen: Qwen2.5 VL 72B Instructqwen88.3%
  2. Qwen/Qwen2-VL-72B-Instructqwen86.1%
  3. OpenGVLab/InternVL2-40Bopengvlab85.1%
  4. Qwen/Qwen2.5-VL-32B-Instructqwen84.7%
  5. NCSOFT/VARCO-VISION-2.0-14Bncsoft84.4%
  6. YannQi/R-4Byannqi83.7%
  7. OpenGVLab/InternVL2_5-8Bopengvlab83.2%
  8. Qwen/Qwen2.5-VL-7B-Instructqwen82.4%
  9. DeepSeek: DeepSeek-VL2deepseek82.3%
  10. OpenGVLab/InternVL3-8Bopengvlab82.3%
  11. Qwen/Qwen2-VL-7B-Instructqwen81.8%
  12. XiaomiMiMo/MiMo-VL-7B-RLxiaomi81.5%

structuralized_imagetext_understanding

mmBench · MMBench · structuralized_imagetext_understanding · percent · Higher is better

  1. Qwen: Qwen2.5 VL 72B Instructqwen88.1%
  2. Qwen/Qwen2-VL-72B-Instructqwen76.2%
  3. OpenGVLab/InternVL2-40Bopengvlab74.3%
  4. DeepSeek: DeepSeek-VL2deepseek66.3%
  5. Google: Gemma 3 12Bgoogle66.3%
  6. Qwen/Qwen2.5-VL-7B-Instructqwen66.3%
  7. rhymes-ai/Ariarhymes-ai66.3%
  8. OpenGVLab/InternVL3-8Bopengvlab65.3%
  9. mistralai/Pixtral-12B-2409mistralai64.4%
  10. moonshotai/Kimi-VL-A3B-Instructmoonshotai64.4%
  11. Qwen/Qwen2.5-VL-3B-Instructqwen64.4%
  12. nvidia/NVLM-D-72Bnvidia61.4%

RR

mmBench · MMBench · RR · percent · Higher is better

  1. OpenGVLab/InternVL2-40Bopengvlab85.3%
  2. OpenGVLab/InternVL3-8Bopengvlab82.5%
  3. Qwen: Qwen2.5 VL 72B Instructqwen81.5%
  4. Qwen/Qwen2-VL-72B-Instructqwen80.1%
  5. Qwen/Qwen2-VL-7B-Instructqwen77.7%
  6. nvidia/NVLM-D-72Bnvidia76.8%
  7. OpenGVLab/InternVL3-2Bopengvlab76.8%
  8. Google: Gemma 3 12Bgoogle76.3%
  9. DeepSeek: DeepSeek-VL2deepseek75.8%
  10. Qwen/Qwen2.5-VL-7B-Instructqwen75.8%
  11. moonshotai/Kimi-VL-A3B-Instructmoonshotai74.9%
  12. rhymes-ai/Ariarhymes-ai74.9%

Models

Loading
EVIDENCE / COMPARISON

Benchmarks

Explore every benchmark

Loading

These charts contain model-linked results. The charts above also include unlinked source profiles. Compare scores within each evaluation. Model quality, usage, transparency and system performance measure different things.

Install TensorTrace