MMBench
Indexed results from MMBench. Each ranking retains its own metric and evaluation conditions.
function_reasoning
mmBench · MMBench · function_reasoning · percent · Higher is better
- Llama-3.2-90B-Vision-Instructmeta-llama97.8%
- DeepSeek: DeepSeek-VL2deepseek96.7%
- OpenGVLab/InternVL2-40Bopengvlab96.7%
- Google: Gemma 3 12Bgoogle95.6%
- moonshotai/Kimi-VL-A3B-Instructmoonshotai95.6%
- Qwen/Qwen2.5-VL-32B-Instructqwen95.6%
- Qwen/Qwen2.5-VL-3B-Instructqwen95.6%
- Qwen: Qwen2.5 VL 72B Instructqwen95.6%
- Qwen/Qwen2.5-VL-7B-Instructqwen95.6%
- mistralai/Pixtral-12B-2409mistralai94.4%
- moonshotai/Kimi-VL-A3B-Thinking-2506moonshotai94.4%
- Qwen/Qwen2-VL-72B-Instructqwen94.4%
image_emotion
mmBench · MMBench · image_emotion · percent · Higher is better
- OpenGVLab/InternVL2-40Bopengvlab87.8%
- Qwen/Qwen2-VL-7B-Instructqwen86.7%
- openbmb/MiniCPM-V-4openbmb85.6%
- OpenGVLab/InternVL2_5-4Bopengvlab85.6%
- OpenGVLab/InternVL3-8Bopengvlab85.6%
- OpenGVLab/InternVL2_5-8Bopengvlab84.4%
- HuggingFaceM4/Idefics3-8B-Llama3huggingfacem483.3%
- Qwen/Qwen2-VL-72B-Instructqwen82.2%
- tiiuae/falcon-11B-vlmtiiuae82.2%
- DeepSeek: DeepSeek-VL2deepseek81.1%
- Google: Gemma 3 27Bgoogle81.1%
- Qwen: Qwen2.5 VL 72B Instructqwen81.1%
FP-C
mmBench · MMBench · FP-C · percent · Higher is better
- Qwen/Qwen2-VL-72B-Instructqwen88%
- Qwen: Qwen2.5 VL 72B Instructqwen85.8%
- YannQi/R-4Byannqi82.4%
- NCSOFT/VARCO-VISION-2.0-14Bncsoft81.6%
- OpenGVLab/InternVL2-40Bopengvlab81.3%
- OpenGVLab/InternVL3-8Bopengvlab79%
- Qwen/Qwen2.5-VL-32B-Instructqwen79%
- DeepSeek: DeepSeek-VL2deepseek78.3%
- OpenGVLab/InternVL2_5-8Bopengvlab78.3%
- Qwen/Qwen2.5-VL-7B-Instructqwen77.2%
- nvidia/NVLM-D-72Bnvidia76.4%
- Qwen/Qwen2-VL-7B-Instructqwen76.4%
future_prediction
mmBench · MMBench · future_prediction · percent · Higher is better
- OpenGVLab/InternVL2-40Bopengvlab80.6%
- Qwen/Qwen2-VL-72B-Instructqwen80.6%
- Qwen: Qwen2.5 VL 72B Instructqwen77.8%
- DeepSeek: DeepSeek-VL2deepseek75%
- OpenGVLab/InternVL3-8Bopengvlab75%
- Google: Gemma 3 12Bgoogle73.6%
- moonshotai/Kimi-VL-A3B-Instructmoonshotai73.6%
- Qwen/Qwen2.5-VL-3B-Instructqwen73.6%
- Qwen/Qwen2.5-VL-7B-Instructqwen73.6%
- OpenGVLab/InternVL3-2Bopengvlab72.2%
- rhymes-ai/Ariarhymes-ai70.8%
- Qwen/Qwen2-VL-7B-Instructqwen69.4%
image_scene
mmBench · MMBench · image_scene · percent · Higher is better
- Qwen: Qwen2.5 VL 72B Instructqwen93.8%
- Qwen/Qwen2.5-VL-32B-Instructqwen92.4%
- OpenGVLab/InternVL2_5-8Bopengvlab91.7%
- Qwen/Qwen2.5-VL-3B-Instructqwen91%
- NCSOFT/VARCO-VISION-2.0-14Bncsoft90.3%
- Qwen/Qwen2.5-VL-7B-Instructqwen90.3%
- tiiuae/falcon-11B-vlmtiiuae89.6%
- vikhyatk/moondream2vikhyatk89.6%
- XiaomiMiMo/MiMo-VL-7B-RLxiaomi89.6%
- microsoft/Phi-3.5-vision-instructmicrosoft88.9%
- microsoft/Phi-3-vision-128k-instructmicrosoft88.9%
- moonshotai/Kimi-VL-A3B-Instructmoonshotai88.9%
object_localization
mmBench · MMBench · object_localization · percent · Higher is better
- Qwen: Qwen2.5 VL 72B Instructqwen90.4%
- OpenGVLab/InternVL2-40Bopengvlab86.5%
- Qwen/Qwen2-VL-72B-Instructqwen84.6%
- Llama-3.2-90B-Vision-Instructmeta-llama82.7%
- moonshotai/Kimi-VL-A3B-Instructmoonshotai80.8%
- OpenGVLab/InternVL3-8Bopengvlab79.8%
- Qwen/Qwen2.5-VL-7B-Instructqwen78.8%
- nvidia/NVLM-D-72Bnvidia76.9%
- rhymes-ai/Ariarhymes-ai76%
- DeepSeek: DeepSeek-VL2deepseek74%
- OpenGVLab/InternVL3-2Bopengvlab73.1%
- ibm-granite/granite-vision-3.2-2bibm-granite72.1%
attribute_comparison
mmBench · MMBench · attribute_comparison · percent · Higher is better
- NCSOFT/VARCO-VISION-2.0-14Bncsoft90.7%
- Qwen: Qwen2.5 VL 72B Instructqwen90.7%
- XiaomiMiMo/MiMo-VL-7B-RLxiaomi90.7%
- YannQi/R-4Byannqi90.7%
- Qwen/Qwen2.5-VL-32B-Instructqwen89.3%
- Qwen/Qwen2-VL-72B-Instructqwen89.3%
- openbmb/MiniCPM-V-4openbmb88%
- OpenGVLab/InternVL2_5-8Bopengvlab88%
- DeepSeek: DeepSeek-VL2deepseek86.7%
- moonshotai/Kimi-VL-A3B-Instructmoonshotai84%
- Qwen/Qwen2.5-VL-7B-Instructqwen84%
- moonshotai/Kimi-VL-A3B-Thinking-2506moonshotai82.7%
object_localization
mmBench · MMBench · object_localization · percent · Higher is better
- Qwen: Qwen2.5 VL 72B Instructqwen91.3%
- Qwen/Qwen2-VL-72B-Instructqwen86.5%
- OpenGVLab/InternVL2-40Bopengvlab81.7%
- Qwen/Qwen2.5-VL-7B-Instructqwen81.7%
- moonshotai/Kimi-VL-A3B-Instructmoonshotai80.8%
- OpenGVLab/InternVL3-8Bopengvlab77.9%
- DeepSeek: DeepSeek-VL2deepseek76.9%
- OpenGVLab/InternVL3-2Bopengvlab75%
- rhymes-ai/Ariarhymes-ai74%
- Llama-3.2-90B-Vision-Instructmeta-llama73.1%
- nvidia/NVLM-D-72Bnvidia72.1%
- Qwen/Qwen2-VL-7B-Instructqwen67.3%
FP-S
mmBench · MMBench · FP-S · percent · Higher is better
- Qwen: Qwen2.5 VL 72B Instructqwen95.1%
- OpenGVLab/InternVL2-40Bopengvlab93.3%
- Qwen/Qwen2-VL-72B-Instructqwen92.3%
- Qwen/Qwen2.5-VL-32B-Instructqwen91%
- NCSOFT/VARCO-VISION-2.0-14Bncsoft90.3%
- OpenGVLab/InternVL2_5-8Bopengvlab90%
- Qwen/Qwen2.5-VL-7B-Instructqwen89.8%
- moonshotai/Kimi-VL-A3B-Instructmoonshotai89.3%
- Qwen/Qwen2-VL-7B-Instructqwen89.3%
- YannQi/R-4Byannqi88.9%
- XiaomiMiMo/MiMo-VL-7B-RLxiaomi88.6%
- DeepSeek: DeepSeek-VL2deepseek88.4%
AR
mmBench · MMBench · AR · percent · Higher is better
- Qwen/Qwen2-VL-72B-Instructqwen89.4%
- Qwen/Qwen2.5-VL-32B-Instructqwen88.6%
- rhymes-ai/Ariarhymes-ai88.6%
- Qwen: Qwen2.5 VL 72B Instructqwen88.2%
- DeepSeek: DeepSeek-VL2deepseek87.3%
- moonshotai/Kimi-VL-A3B-Instructmoonshotai87.3%
- NCSOFT/VARCO-VISION-2.0-14Bncsoft87.3%
- OpenGVLab/InternVL2-40Bopengvlab87.3%
- Google: Gemma 3 27Bgoogle86.9%
- Llama-3.2-90B-Vision-Instructmeta-llama86.5%
- openbmb/MiniCPM-V-4openbmb85.7%
- YannQi/R-4Byannqi85.7%
image_quality
mmBench · MMBench · image_quality · percent · Higher is better
- Qwen: Qwen2.5 VL 72B Instructqwen88.2%
- DeepSeek: DeepSeek-VL2deepseek85.9%
- OpenGVLab/InternVL2-40Bopengvlab84.7%
- Qwen/Qwen2-VL-72B-Instructqwen83.5%
- Qwen/Qwen2.5-VL-7B-Instructqwen82.4%
- OpenGVLab/InternVL3-8Bopengvlab81.2%
- moonshotai/Kimi-VL-A3B-Instructmoonshotai80%
- nvidia/NVLM-D-72Bnvidia78.8%
- rhymes-ai/Ariarhymes-ai78.8%
- ibm-granite/granite-vision-3.2-2bibm-granite76.5%
- Llama-3.2-90B-Vision-Instructmeta-llama76.5%
- OpenGVLab/InternVL3-2Bopengvlab76.5%
identity_reasoning
mmBench · MMBench · identity_reasoning · percent · Higher is better
- Google: Gemma 3 27Bgoogle100%
- moonshotai/Kimi-VL-A3B-Thinking-2506moonshotai98.7%
- NCSOFT/VARCO-VISION-2.0-14Bncsoft98.7%
- nvidia/NVLM-D-72Bnvidia98.7%
- OpenGVLab/InternVL2_5-8Bopengvlab98.7%
- OpenGVLab/InternVL3-8Bopengvlab98.7%
- Qwen/Qwen2.5-VL-32B-Instructqwen98.7%
- Qwen/Qwen2-VL-7B-Instructqwen98.7%
- Google: Gemma 3 12Bgoogle97.4%
- HuggingFaceM4/Idefics3-8B-Llama3huggingfacem497.4%
- Qwen/Qwen2.5-VL-7B-Instructqwen97.4%
- Qwen/Qwen2-VL-2B-Instructqwen97.4%
Models
LoadingBenchmarks
Explore every benchmark
LoadingThese 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.