MMBench

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

physical_relation

mmBench · MMBench · physical_relation · percent · Higher is better

  1. OpenGVLab/InternVL2-40Bopengvlab75%
  2. moonshotai/Kimi-VL-A3B-Instructmoonshotai73.1%
  3. OpenGVLab/InternVL3-8Bopengvlab73.1%
  4. DeepSeek: DeepSeek-VL2deepseek71.2%
  5. nvidia/NVLM-D-72Bnvidia71.2%
  6. rhymes-ai/Ariarhymes-ai69.2%
  7. OpenGVLab/InternVL3-2Bopengvlab67.3%
  8. allenai/Molmo-7B-O-0924allenai63.5%
  9. h2oai/h2ovl-mississippi-2bh2oai61.5%
  10. ibm-granite/granite-vision-3.2-2bibm-granite61.5%
  11. Llama-3.2-90B-Vision-Instructmeta-llama61.5%
  12. mistralai/Pixtral-12B-2409mistralai59.6%

celebrity_recognition

mmBench · MMBench · celebrity_recognition · percent · Higher is better

  1. Qwen: Qwen2.5 VL 72B Instructqwen96.6%
  2. Qwen/Qwen2-VL-72B-Instructqwen96.6%
  3. Qwen/Qwen2-VL-7B-Instructqwen93.2%
  4. OpenGVLab/InternVL2-40Bopengvlab91.8%
  5. Qwen/Qwen2.5-VL-3B-Instructqwen91.8%
  6. Qwen/Qwen2.5-VL-7B-Instructqwen91.2%
  7. Google: Gemma 3 27Bgoogle90.5%
  8. Llama-3.2-90B-Vision-Instructmeta-llama89.8%
  9. XiaomiMiMo/MiMo-VL-7B-RLxiaomi89.8%
  10. HuggingFaceM4/idefics2-8bhuggingfacem489.1%
  11. Qwen/Qwen2.5-VL-32B-Instructqwen89.1%
  12. Xkev/Llama-3.2V-11B-cotxkev89.1%

image_emotion

mmBench · MMBench · image_emotion · percent · Higher is better

  1. Qwen/Qwen2-VL-7B-Instructqwen83.1%
  2. DeepSeek: DeepSeek-VL2deepseek80.7%
  3. OpenGVLab/InternVL2-40Bopengvlab80.7%
  4. OpenGVLab/InternVL3-8Bopengvlab80.7%
  5. microsoft/Phi-3.5-vision-instructmicrosoft78.3%
  6. Qwen/Qwen2-VL-2B-Instructqwen78.3%
  7. OpenGVLab/InternVL3-2Bopengvlab77.1%
  8. Qwen: Qwen2.5 VL 72B Instructqwen77.1%
  9. Qwen/Qwen2.5-VL-7B-Instructqwen77.1%
  10. Qwen/Qwen2-VL-72B-Instructqwen77.1%
  11. tiiuae/falcon-11B-vlmtiiuae77.1%
  12. HuggingFaceM4/Idefics3-8B-Llama3huggingfacem475.9%

attribute_comparison

mmBench · MMBench · attribute_comparison · percent · Higher is better

  1. Qwen/Qwen2-VL-72B-Instructqwen90.7%
  2. YannQi/R-4Byannqi90.7%
  3. NCSOFT/VARCO-VISION-2.0-14Bncsoft88%
  4. Qwen: Qwen2.5 VL 72B Instructqwen86.7%
  5. openbmb/MiniCPM-V-4openbmb84%
  6. OpenGVLab/InternVL2_5-8Bopengvlab84%
  7. Qwen/Qwen2.5-VL-32B-Instructqwen82.7%
  8. DeepSeek: DeepSeek-VL2deepseek81.3%
  9. OpenGVLab/InternVL3-8Bopengvlab78.7%
  10. XiaomiMiMo/MiMo-VL-7B-RLxiaomi78.7%
  11. OpenGVLab/InternVL2-40Bopengvlab76%
  12. Qwen/Qwen2.5-VL-7B-Instructqwen76%

RR

mmBench · MMBench · RR · percent · Higher is better

  1. OpenGVLab/InternVL2-40Bopengvlab89.1%
  2. OpenGVLab/InternVL3-2Bopengvlab84.4%
  3. Qwen: Qwen2.5 VL 72B Instructqwen82.9%
  4. OpenGVLab/InternVL3-8Bopengvlab82.5%
  5. DeepSeek: DeepSeek-VL2deepseek82%
  6. Qwen/Qwen2-VL-72B-Instructqwen82%
  7. nvidia/NVLM-D-72Bnvidia80.6%
  8. moonshotai/Kimi-VL-A3B-Instructmoonshotai79.6%
  9. rhymes-ai/Ariarhymes-ai79.6%
  10. Qwen/Qwen2.5-VL-7B-Instructqwen78.2%
  11. Llama-3.2-90B-Vision-Instructmeta-llama77.7%
  12. microsoft/Phi-3.5-vision-instructmicrosoft74.9%

spatial_relationship

mmBench · MMBench · spatial_relationship · percent · Higher is better

  1. Qwen: Qwen2.5 VL 72B Instructqwen79%
  2. Qwen/Qwen2-VL-72B-Instructqwen77.8%
  3. nvidia/NVLM-D-72Bnvidia63%
  4. OpenGVLab/InternVL2-40Bopengvlab63%
  5. Qwen/Qwen2.5-VL-7B-Instructqwen59.3%
  6. DeepSeek: DeepSeek-VL2deepseek55.6%
  7. OpenGVLab/InternVL3-8Bopengvlab55.6%
  8. moonshotai/Kimi-VL-A3B-Instructmoonshotai53.1%
  9. Qwen/Qwen2-VL-7B-Instructqwen51.9%
  10. Llama-3.2-90B-Vision-Instructmeta-llama50.6%
  11. mistralai/Pixtral-12B-2409mistralai50.6%
  12. rhymes-ai/Ariarhymes-ai43.2%

identity_reasoning

mmBench · MMBench · identity_reasoning · percent · Higher is better

  1. Qwen/Qwen2-VL-72B-Instructqwen100%
  2. Google: Gemma 3 27Bgoogle98.7%
  3. moonshotai/Kimi-VL-A3B-Thinking-2506moonshotai98.7%
  4. OpenGVLab/InternVL3-8Bopengvlab98.7%
  5. Xkev/Llama-3.2V-11B-cotxkev98.7%
  6. DeepSeek: DeepSeek-VL2deepseek97.4%
  7. HuggingFaceM4/idefics2-8bhuggingfacem497.4%
  8. microsoft/Phi-3.5-vision-instructmicrosoft97.4%
  9. Qwen/Qwen2.5-VL-3B-Instructqwen97.4%
  10. Qwen: Qwen2.5 VL 72B Instructqwen97.4%
  11. Qwen/Qwen2-VL-2B-Instructqwen97.4%
  12. vikhyatk/moondream2vikhyatk97.4%

celebrity_recognition

mmBench · MMBench · celebrity_recognition · percent · Higher is better

  1. Qwen: Qwen2.5 VL 72B Instructqwen97.5%
  2. Qwen/Qwen2-VL-72B-Instructqwen96.6%
  3. OpenGVLab/InternVL2-40Bopengvlab94.1%
  4. Qwen/Qwen2.5-VL-3B-Instructqwen93.2%
  5. Qwen/Qwen2.5-VL-7B-Instructqwen92.4%
  6. Llama-3.2-90B-Vision-Instructmeta-llama90.7%
  7. moonshotai/Kimi-VL-A3B-Instructmoonshotai90.7%
  8. Qwen/Qwen2-VL-7B-Instructqwen90.7%
  9. nvidia/NVLM-D-72Bnvidia88.1%
  10. rhymes-ai/Ariarhymes-ai86.4%
  11. Qwen/Qwen2-VL-2B-Instructqwen85.6%
  12. Google: Gemma 3 12Bgoogle84.7%

spatial_relationship

mmBench · MMBench · spatial_relationship · percent · Higher is better

  1. Qwen: Qwen2.5 VL 72B Instructqwen84%
  2. Qwen/Qwen2-VL-72B-Instructqwen80%
  3. Qwen/Qwen2.5-VL-32B-Instructqwen73.3%
  4. OpenGVLab/InternVL2-40Bopengvlab69.3%
  5. YannQi/R-4Byannqi68%
  6. NCSOFT/VARCO-VISION-2.0-14Bncsoft66.7%
  7. nvidia/NVLM-D-72Bnvidia66.7%
  8. DeepSeek: DeepSeek-VL2deepseek62.7%
  9. Qwen/Qwen2.5-VL-7B-Instructqwen62.7%
  10. OpenGVLab/InternVL3-8Bopengvlab58.7%
  11. XiaomiMiMo/MiMo-VL-7B-RLxiaomi57.3%
  12. moonshotai/Kimi-VL-A3B-Instructmoonshotai56%

future_prediction

mmBench · MMBench · future_prediction · percent · Higher is better

  1. Qwen/Qwen2-VL-72B-Instructqwen85.3%
  2. OpenGVLab/InternVL2-40Bopengvlab84%
  3. NCSOFT/VARCO-VISION-2.0-14Bncsoft82.7%
  4. Google: Gemma 3 27Bgoogle81.3%
  5. Qwen: Qwen2.5 VL 72B Instructqwen81.3%
  6. DeepSeek: DeepSeek-VL2deepseek80%
  7. moonshotai/Kimi-VL-A3B-Instructmoonshotai80%
  8. Qwen/Qwen2.5-VL-7B-Instructqwen80%
  9. Google: Gemma 3 12Bgoogle78.7%
  10. OpenGVLab/InternVL2_5-4Bopengvlab78.7%
  11. OpenGVLab/InternVL2_5-8Bopengvlab78.7%
  12. Qwen/Qwen2.5-VL-32B-Instructqwen77.3%

image_topic

mmBench · MMBench · image_topic · percent · Higher is better

  1. HuggingFaceM4/Idefics3-8B-Llama3huggingfacem497.6%
  2. OpenGVLab/InternVL2-40Bopengvlab97.6%
  3. OpenGVLab/InternVL3-8Bopengvlab97.6%
  4. Google: Gemma 3 12Bgoogle96.5%
  5. Google: Gemma 3 4Bgoogle96.5%
  6. Llama-3.2-90B-Vision-Instructmeta-llama96.5%
  7. mistralai/Pixtral-12B-2409mistralai96.5%
  8. moonshotai/Kimi-VL-A3B-Instructmoonshotai96.5%
  9. OpenGVLab/InternVL3-2Bopengvlab96.5%
  10. Qwen/Qwen2.5-VL-3B-Instructqwen96.5%
  11. Qwen: Qwen2.5 VL 72B Instructqwen96.5%
  12. Qwen/Qwen2-VL-72B-Instructqwen96.5%

FP-S

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

  1. Qwen: Qwen2.5 VL 72B Instructqwen96.3%
  2. Qwen/Qwen2-VL-72B-Instructqwen92.8%
  3. OpenGVLab/InternVL2-40Bopengvlab91.6%
  4. Qwen/Qwen2.5-VL-32B-Instructqwen90.3%
  5. Qwen/Qwen2.5-VL-7B-Instructqwen89.6%
  6. moonshotai/Kimi-VL-A3B-Instructmoonshotai89.3%
  7. OpenGVLab/InternVL2_5-8Bopengvlab88.9%
  8. nvidia/NVLM-D-72Bnvidia87.7%
  9. XiaomiMiMo/MiMo-VL-7B-RLxiaomi87%
  10. OpenGVLab/InternVL3-8Bopengvlab86.8%
  11. NCSOFT/VARCO-VISION-2.0-14Bncsoft86.5%
  12. DeepSeek: DeepSeek-VL2deepseek86.1%

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