MMBench
Indexed results from MMBench. Each ranking retains its own metric and evaluation conditions.
nature_relation
mmBench · MMBench · nature_relation · percent · Higher is better
- NCSOFT/VARCO-VISION-2.0-14Bncsoft93.5%
- OpenGVLab/InternVL2-40Bopengvlab92.4%
- Qwen: Qwen2.5 VL 72B Instructqwen91.3%
- YannQi/R-4Byannqi90.2%
- Google: Gemma 3 12Bgoogle89.1%
- Google: Gemma 3 27Bgoogle89.1%
- OpenGVLab/InternVL3-8Bopengvlab89.1%
- Qwen/Qwen2-VL-72B-Instructqwen89.1%
- Llama-3.2-90B-Vision-Instructmeta-llama88%
- moonshotai/Kimi-VL-A3B-Thinking-2506moonshotai88%
- Qwen/Qwen2.5-VL-32B-Instructqwen88%
- OpenGVLab/InternVL2_5-8Bopengvlab87%
object_localization
mmBench · MMBench · object_localization · percent · Higher is better
- Qwen/Qwen2.5-VL-32B-Instructqwen90.5%
- Qwen: Qwen2.5 VL 72B Instructqwen89.5%
- NCSOFT/VARCO-VISION-2.0-14Bncsoft88.6%
- OpenGVLab/InternVL2-40Bopengvlab88.6%
- OpenGVLab/InternVL2_5-8Bopengvlab85.7%
- Qwen/Qwen2-VL-72B-Instructqwen85.7%
- moonshotai/Kimi-VL-A3B-Instructmoonshotai83.8%
- XiaomiMiMo/MiMo-VL-7B-RLxiaomi82.9%
- YannQi/R-4Byannqi82.9%
- OpenGVLab/InternVL3-8Bopengvlab81.9%
- Qwen/Qwen2.5-VL-7B-Instructqwen81.9%
- nvidia/NVLM-D-72Bnvidia80%
action_recognition
mmBench · MMBench · action_recognition · percent · Higher is better
- OpenGVLab/InternVL3-8Bopengvlab96.6%
- Qwen/Qwen2-VL-72B-Instructqwen95.5%
- DeepSeek: DeepSeek-VL2deepseek94.3%
- h2oai/h2ovl-mississippi-2bh2oai94.3%
- OpenGVLab/InternVL2-1Bopengvlab94.3%
- OpenGVLab/InternVL3-2Bopengvlab94.3%
- Qwen/Qwen2.5-VL-7B-Instructqwen94.3%
- Qwen/Qwen2-VL-2B-Instructqwen94.3%
- Qwen/Qwen2-VL-7B-Instructqwen94.3%
- tiiuae/falcon-11B-vlmtiiuae94.3%
- moonshotai/Kimi-VL-A3B-Instructmoonshotai93.2%
- nvidia/NVLM-D-72Bnvidia93.2%
image_scene
mmBench · MMBench · image_scene · percent · Higher is better
- OpenGVLab/InternVL2_5-8Bopengvlab92.4%
- Qwen/Qwen2.5-VL-3B-Instructqwen92.4%
- Qwen/Qwen2.5-VL-7B-Instructqwen92.4%
- OpenGVLab/InternVL2_5-4Bopengvlab91.7%
- Qwen: Qwen2.5 VL 72B Instructqwen91.7%
- NCSOFT/VARCO-VISION-2.0-14Bncsoft90.3%
- openbmb/MiniCPM-V-4openbmb90.3%
- DeepSeek: DeepSeek-VL2deepseek89.6%
- moonshotai/Kimi-VL-A3B-Instructmoonshotai89.6%
- OpenGVLab/InternVL2-40Bopengvlab89.6%
- Qwen/Qwen2.5-VL-32B-Instructqwen88.9%
- tiiuae/falcon-11B-vlmtiiuae88.9%
nature_relation
mmBench · MMBench · nature_relation · percent · Higher is better
- OpenGVLab/InternVL2-40Bopengvlab95.1%
- Qwen: Qwen2.5 VL 72B Instructqwen91.4%
- OpenGVLab/InternVL3-2Bopengvlab87.7%
- Qwen/Qwen2.5-VL-7B-Instructqwen87.7%
- Qwen/Qwen2-VL-72B-Instructqwen86.4%
- DeepSeek: DeepSeek-VL2deepseek85.2%
- OpenGVLab/InternVL3-8Bopengvlab85.2%
- rhymes-ai/Ariarhymes-ai84%
- nvidia/NVLM-D-72Bnvidia82.7%
- Google: Gemma 3 12Bgoogle81.5%
- Llama-3.2-90B-Vision-Instructmeta-llama79%
- microsoft/Phi-3.5-vision-instructmicrosoft79%
image_topic
mmBench · MMBench · image_topic · percent · Higher is better
- HuggingFaceM4/idefics2-8bhuggingfacem498.9%
- mistralai/Pixtral-12B-2409mistralai98.9%
- Qwen/Qwen2.5-VL-32B-Instructqwen98.9%
- Qwen: Qwen2.5 VL 72B Instructqwen98.9%
- Qwen/Qwen2-VL-2B-Instructqwen98.9%
- rhymes-ai/Ariarhymes-ai98.9%
- Google: Gemma 3 12Bgoogle97.8%
- Google: Gemma 3 27Bgoogle97.8%
- ibm-granite/granite-vision-3.2-2bibm-granite97.8%
- Qwen/Qwen2.5-VL-3B-Instructqwen97.8%
- Qwen/Qwen2.5-VL-7B-Instructqwen97.8%
- tiiuae/falcon-11B-vlmtiiuae97.8%
ocr
mmBench · MMBench · ocr · percent · Higher is better
- Qwen: Qwen2.5 VL 72B Instructqwen100%
- nvidia/NVLM-D-72Bnvidia98.7%
- moonshotai/Kimi-VL-A3B-Instructmoonshotai97.4%
- Qwen/Qwen2-VL-72B-Instructqwen97.4%
- Llama-3.2-90B-Vision-Instructmeta-llama96.1%
- OpenGVLab/InternVL2-40Bopengvlab96.1%
- OpenGVLab/InternVL3-8Bopengvlab96.1%
- Qwen/Qwen2.5-VL-7B-Instructqwen96.1%
- Qwen/Qwen2-VL-7B-Instructqwen96.1%
- rhymes-ai/Ariarhymes-ai96.1%
- Google: Gemma 3 4Bgoogle94.8%
- Qwen/Qwen2.5-VL-3B-Instructqwen94.8%
image_style
mmBench · MMBench · image_style · percent · Higher is better
- Qwen: Qwen2.5 VL 72B Instructqwen83.7%
- Qwen/Qwen2.5-VL-7B-Instructqwen82.6%
- Qwen/Qwen2-VL-7B-Instructqwen82.6%
- Google: Gemma 3 27Bgoogle81.5%
- OpenGVLab/InternVL2-40Bopengvlab81.5%
- OpenGVLab/InternVL2_5-8Bopengvlab81.5%
- Qwen/Qwen2-VL-72B-Instructqwen80.4%
- Google: Gemma 3 12Bgoogle79.3%
- moonshotai/Kimi-VL-A3B-Instructmoonshotai79.3%
- DeepSeek: DeepSeek-VL2deepseek78.3%
- OpenGVLab/InternVL2_5-4Bopengvlab78.3%
- Qwen/Qwen2.5-VL-32B-Instructqwen78.3%
physical_property_reasoning
mmBench · MMBench · physical_property_reasoning · percent · Higher is better
- DeepSeek: DeepSeek-VL2deepseek76%
- OpenGVLab/InternVL2-40Bopengvlab73%
- rhymes-ai/Ariarhymes-ai73%
- HuggingFaceM4/Idefics3-8B-Llama3huggingfacem471%
- Qwen/Qwen2-VL-72B-Instructqwen70%
- moonshotai/Kimi-VL-A3B-Instructmoonshotai69%
- OpenGVLab/InternVL2-1Bopengvlab68%
- OpenGVLab/InternVL3-8Bopengvlab68%
- Qwen: Qwen2.5 VL 72B Instructqwen66%
- 5CD-AI/Vintern-3B-beta5cd-ai65%
- Qwen/Qwen2.5-VL-7B-Instructqwen65%
- tiiuae/falcon-11B-vlmtiiuae65%
spatial_relationship
mmBench · MMBench · spatial_relationship · percent · Higher is better
- Qwen/Qwen2-VL-72B-Instructqwen81.5%
- Qwen: Qwen2.5 VL 72B Instructqwen80.2%
- nvidia/NVLM-D-72Bnvidia65.4%
- OpenGVLab/InternVL2-40Bopengvlab65.4%
- Qwen/Qwen2.5-VL-7B-Instructqwen61.7%
- OpenGVLab/InternVL3-8Bopengvlab56.8%
- Llama-3.2-90B-Vision-Instructmeta-llama55.6%
- DeepSeek: DeepSeek-VL2deepseek49.4%
- mistralai/Pixtral-12B-2409mistralai48.1%
- rhymes-ai/Ariarhymes-ai48.1%
- Google: Gemma 3 12Bgoogle46.9%
- HuggingFaceM4/Idefics3-8B-Llama3huggingfacem442%
AR
mmBench · MMBench · AR · percent · Higher is better
- DeepSeek: DeepSeek-VL2deepseek87.5%
- OpenGVLab/InternVL2-40Bopengvlab86.5%
- Qwen/Qwen2-VL-72B-Instructqwen85.4%
- rhymes-ai/Ariarhymes-ai85.4%
- moonshotai/Kimi-VL-A3B-Instructmoonshotai85.1%
- Qwen: Qwen2.5 VL 72B Instructqwen84%
- HuggingFaceM4/Idefics3-8B-Llama3huggingfacem483.7%
- OpenGVLab/InternVL3-8Bopengvlab83.3%
- Llama-3.2-90B-Vision-Instructmeta-llama82.6%
- OpenGVLab/InternVL3-2Bopengvlab82.6%
- Qwen/Qwen2.5-VL-3B-Instructqwen82.6%
- Qwen/Qwen2.5-VL-7B-Instructqwen82.6%
future_prediction
mmBench · MMBench · future_prediction · percent · Higher is better
- Qwen/Qwen2-VL-72B-Instructqwen79.2%
- OpenGVLab/InternVL2-40Bopengvlab77.8%
- Qwen: Qwen2.5 VL 72B Instructqwen77.8%
- nvidia/NVLM-D-72Bnvidia76.4%
- Qwen/Qwen2.5-VL-7B-Instructqwen76.4%
- DeepSeek: DeepSeek-VL2deepseek75%
- Google: Gemma 3 12Bgoogle75%
- OpenGVLab/InternVL3-8Bopengvlab75%
- Google: Gemma 3 4Bgoogle73.6%
- Llama-3.2-90B-Vision-Instructmeta-llama73.6%
- rhymes-ai/Ariarhymes-ai73.6%
- vikhyatk/moondream2vikhyatk72.2%
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.