MMBench

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

structuralized_imagetext_understanding

mmBench · MMBench · structuralized_imagetext_understanding · percent · Higher is better

  1. Qwen: Qwen2.5 VL 72B Instructqwen91.1%
  2. OpenGVLab/InternVL2-40Bopengvlab79.2%
  3. Qwen/Qwen2-VL-72B-Instructqwen78.2%
  4. OpenGVLab/InternVL3-8Bopengvlab72.3%
  5. rhymes-ai/Ariarhymes-ai72.3%
  6. mistralai/Pixtral-12B-2409mistralai71.3%
  7. Qwen/Qwen2.5-VL-3B-Instructqwen71.3%
  8. moonshotai/Kimi-VL-A3B-Instructmoonshotai69.3%
  9. Qwen/Qwen2.5-VL-7B-Instructqwen69.3%
  10. DeepSeek: DeepSeek-VL2deepseek68.3%
  11. nvidia/NVLM-D-72Bnvidia68.3%
  12. Google: Gemma 3 12Bgoogle67.3%

spatial_relationship

mmBench · MMBench · spatial_relationship · percent · Higher is better

  1. Qwen: Qwen2.5 VL 72B Instructqwen84%
  2. Qwen/Qwen2-VL-72B-Instructqwen84%
  3. Qwen/Qwen2.5-VL-32B-Instructqwen73.3%
  4. YannQi/R-4Byannqi72%
  5. nvidia/NVLM-D-72Bnvidia69.3%
  6. XiaomiMiMo/MiMo-VL-7B-RLxiaomi69.3%
  7. Llama-3.2-90B-Vision-Instructmeta-llama68%
  8. OpenGVLab/InternVL2-40Bopengvlab66.7%
  9. Qwen/Qwen2.5-VL-7B-Instructqwen66.7%
  10. moonshotai/Kimi-VL-A3B-Thinking-2506moonshotai64%
  11. NCSOFT/VARCO-VISION-2.0-14Bncsoft62.7%
  12. OpenGVLab/InternVL3-8Bopengvlab61.3%

attribute_recognition

mmBench · MMBench · attribute_recognition · percent · Higher is better

  1. OpenGVLab/InternVL2-40Bopengvlab98%
  2. Qwen: Qwen2.5 VL 72B Instructqwen97%
  3. nvidia/NVLM-D-72Bnvidia94.9%
  4. Qwen/Qwen2-VL-72B-Instructqwen93.9%
  5. OpenGVLab/InternVL3-8Bopengvlab92.9%
  6. moonshotai/Kimi-VL-A3B-Instructmoonshotai90.9%
  7. rhymes-ai/Ariarhymes-ai90.9%
  8. DeepSeek: DeepSeek-VL2deepseek89.9%
  9. OpenGVLab/InternVL3-2Bopengvlab88.9%
  10. Qwen/Qwen2-VL-7B-Instructqwen88.9%
  11. Qwen/Qwen2.5-VL-3B-Instructqwen86.9%
  12. Qwen/Qwen2.5-VL-7B-Instructqwen86.9%

object_localization

mmBench · MMBench · object_localization · percent · Higher is better

  1. Qwen: Qwen2.5 VL 72B Instructqwen92.4%
  2. Qwen/Qwen2-VL-72B-Instructqwen86.7%
  3. Qwen/Qwen2.5-VL-32B-Instructqwen84.8%
  4. moonshotai/Kimi-VL-A3B-Instructmoonshotai83.8%
  5. OpenGVLab/InternVL2-40Bopengvlab83.8%
  6. OpenGVLab/InternVL2_5-8Bopengvlab83.8%
  7. NCSOFT/VARCO-VISION-2.0-14Bncsoft82.9%
  8. Qwen/Qwen2.5-VL-7B-Instructqwen82.9%
  9. YannQi/R-4Byannqi81%
  10. OpenGVLab/InternVL3-8Bopengvlab79%
  11. DeepSeek: DeepSeek-VL2deepseek78.1%
  12. XiaomiMiMo/MiMo-VL-7B-RLxiaomi78.1%

FP-S

mmBench · MMBench · FP-S · percent · Higher is better

  1. Qwen: Qwen2.5 VL 72B Instructqwen96.2%
  2. Qwen/Qwen2-VL-72B-Instructqwen93.5%
  3. OpenGVLab/InternVL2-40Bopengvlab92.2%
  4. moonshotai/Kimi-VL-A3B-Instructmoonshotai89.4%
  5. Qwen/Qwen2.5-VL-7B-Instructqwen88.9%
  6. nvidia/NVLM-D-72Bnvidia87.7%
  7. OpenGVLab/InternVL3-8Bopengvlab86.7%
  8. Llama-3.2-90B-Vision-Instructmeta-llama86.2%
  9. rhymes-ai/Ariarhymes-ai86.2%
  10. DeepSeek: DeepSeek-VL2deepseek85.9%
  11. Qwen/Qwen2-VL-7B-Instructqwen85.2%
  12. Qwen/Qwen2.5-VL-3B-Instructqwen83.9%

calligraphy_painting

mmBench · MMBench · calligraphy_painting · percent · Higher is better

  1. OpenGVLab/InternVL2-1Bopengvlab91.2%
  2. OpenGVLab/InternVL3-8Bopengvlab91.2%
  3. OpenGVLab/InternVL3-2Bopengvlab89.5%
  4. OpenGVLab/InternVL2-40Bopengvlab87.7%
  5. moonshotai/Kimi-VL-A3B-Instructmoonshotai71.9%
  6. Qwen: Qwen2.5 VL 72B Instructqwen71.9%
  7. Qwen/Qwen2-VL-72B-Instructqwen70.2%
  8. Qwen/Qwen2-VL-7B-Instructqwen70.2%
  9. DeepSeek: DeepSeek-VL2deepseek66.7%
  10. Qwen/Qwen2.5-VL-3B-Instructqwen63.2%
  11. nvidia/NVLM-D-72Bnvidia59.6%
  12. Google: Gemma 3 12Bgoogle57.9%

identity_reasoning

mmBench · MMBench · identity_reasoning · percent · Higher is better

  1. Qwen/Qwen2-VL-72B-Instructqwen100%
  2. OpenGVLab/InternVL3-8Bopengvlab98.8%
  3. DeepSeek: DeepSeek-VL2deepseek97.6%
  4. microsoft/Phi-3.5-vision-instructmicrosoft97.6%
  5. moonshotai/Kimi-VL-A3B-Instructmoonshotai97.6%
  6. OpenGVLab/InternVL3-2Bopengvlab97.6%
  7. Qwen/Qwen2.5-VL-3B-Instructqwen97.6%
  8. Qwen: Qwen2.5 VL 72B Instructqwen97.6%
  9. vikhyatk/moondream2vikhyatk97.6%
  10. ibm-granite/granite-vision-3.2-2bibm-granite96.3%
  11. Qwen/Qwen2.5-VL-7B-Instructqwen96.3%
  12. tiiuae/falcon-11B-vlmtiiuae96.3%

AR

mmBench · MMBench · AR · percent · Higher is better

  1. OpenGVLab/InternVL2-40Bopengvlab87.5%
  2. Qwen/Qwen2-VL-72B-Instructqwen85.1%
  3. Qwen: Qwen2.5 VL 72B Instructqwen84.4%
  4. DeepSeek: DeepSeek-VL2deepseek84%
  5. OpenGVLab/InternVL3-8Bopengvlab84%
  6. nvidia/NVLM-D-72Bnvidia83.7%
  7. moonshotai/Kimi-VL-A3B-Instructmoonshotai83.3%
  8. Qwen/Qwen2.5-VL-7B-Instructqwen83.3%
  9. OpenGVLab/InternVL3-2Bopengvlab83%
  10. rhymes-ai/Ariarhymes-ai83%
  11. Qwen/Qwen2-VL-7B-Instructqwen82.6%
  12. Google: Gemma 3 12Bgoogle81.6%

image_quality

mmBench · MMBench · image_quality · percent · Higher is better

  1. Qwen: Qwen2.5 VL 72B Instructqwen71%
  2. XiaomiMiMo/MiMo-VL-7B-RLxiaomi68.5%
  3. Qwen/Qwen2.5-VL-32B-Instructqwen63.7%
  4. Qwen/Qwen2-VL-72B-Instructqwen63.7%
  5. DeepSeek: DeepSeek-VL2deepseek62.9%
  6. moonshotai/Kimi-VL-A3B-Thinking-2506moonshotai62.1%
  7. OpenGVLab/InternVL2-40Bopengvlab61.3%
  8. NCSOFT/VARCO-VISION-2.0-14Bncsoft60.5%
  9. Qwen/Qwen2.5-VL-7B-Instructqwen60.5%
  10. YannQi/R-4Byannqi60.5%
  11. moonshotai/Kimi-VL-A3B-Instructmoonshotai59.7%
  12. Qwen/Qwen2-VL-7B-Instructqwen59.7%

traditional_show

mmBench · MMBench · traditional_show · percent · Higher is better

  1. OpenGVLab/InternVL3-8Bopengvlab90.9%
  2. OpenGVLab/InternVL2-40Bopengvlab84.8%
  3. OpenGVLab/InternVL2-1Bopengvlab83.3%
  4. OpenGVLab/InternVL3-2Bopengvlab75.8%
  5. Qwen: Qwen2.5 VL 72B Instructqwen71.2%
  6. Qwen/Qwen2-VL-72B-Instructqwen69.7%
  7. Qwen/Qwen2-VL-7B-Instructqwen63.6%
  8. moonshotai/Kimi-VL-A3B-Instructmoonshotai62.1%
  9. DeepSeek: DeepSeek-VL2deepseek60.6%
  10. nvidia/NVLM-D-72Bnvidia57.6%
  11. 5CD-AI/Vintern-3B-beta5cd-ai56.1%
  12. Qwen/Qwen2.5-VL-3B-Instructqwen56.1%

scenery_building

mmBench · MMBench · scenery_building · percent · Higher is better

  1. Qwen: Qwen2.5 VL 72B Instructqwen94.7%
  2. moonshotai/Kimi-VL-A3B-Instructmoonshotai90.5%
  3. Qwen/Qwen2-VL-72B-Instructqwen88.4%
  4. OpenGVLab/InternVL3-2Bopengvlab83.2%
  5. Qwen/Qwen2.5-VL-3B-Instructqwen83.2%
  6. Qwen/Qwen2.5-VL-7B-Instructqwen83.2%
  7. OpenGVLab/InternVL2-1Bopengvlab80%
  8. OpenGVLab/InternVL2-40Bopengvlab80%
  9. Qwen/Qwen2-VL-7B-Instructqwen77.9%
  10. OpenGVLab/InternVL3-8Bopengvlab76.8%
  11. Llama-3.2-90B-Vision-Instructmeta-llama70.5%
  12. Qwen/Qwen2-VL-2B-Instructqwen62.1%

Overall

mmBench · MMBench · Overall · percent · Higher is better

  1. OpenGVLab/InternVL2-40Bopengvlab80.2%
  2. OpenGVLab/InternVL3-8Bopengvlab77.8%
  3. OpenGVLab/InternVL2-1Bopengvlab76.3%
  4. OpenGVLab/InternVL3-2Bopengvlab75.3%
  5. Qwen: Qwen2.5 VL 72B Instructqwen73.9%
  6. moonshotai/Kimi-VL-A3B-Instructmoonshotai72%
  7. Qwen/Qwen2-VL-72B-Instructqwen69.8%
  8. Qwen/Qwen2-VL-7B-Instructqwen65.7%
  9. Qwen/Qwen2.5-VL-3B-Instructqwen62%
  10. Qwen/Qwen2.5-VL-7B-Instructqwen59.2%
  11. nvidia/NVLM-D-72Bnvidia59%
  12. DeepSeek: DeepSeek-VL2deepseek55.7%

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