List your company
ResearchFormerly WizWand

Multimodal Understanding on MMBench (dev)

93.2MMB ScoreGemini 3.1 Pro
-3.1164821.8887646.89471.89924Mar 3, 2024Jul 17, 2024Dec 1, 2024Apr 17, 2025Aug 31, 2025Jan 15, 2026Jun 1, 2026
Updated 4mo ago

Evaluation results

MethodLinks
Gemini 3.1 ProAccess=Closed Model2026.0693.2——
VanillaModels=BAGEL-7B, Avg. Tokens=100%2026.0588.5—100
VSCANModels=BAGEL-7B, Avg. Tokens=50%2026.0588.5—94.4
G2TRModels=BAGEL-7B, Avg. Tokens=50%2026.0588.5—99
IVC-PruneModels=BAGEL-7B, Avg. Tokens=50%2026.0588.4—98.6
Qwen3-VL 32BModel Size=32B2026.0687.8——
Cosmos 3 SuperModel Size=32B2026.0687.4——
Gemma-4 31BModel Size=31B2026.0686.7——
Cosmos-Reason2 32BModel Size=32B2026.0686.3——
W-FastVModels=BAGEL-7B, Avg. Tokens=54%2026.0585.9—95.6
FastVModels=BAGEL-7B, Avg. Tokens=54%2026.0585.7—95.8
RynnBrain 8BModel Size=8B2026.0685.5——
Qwen3-VL 8BModel Size=8B2026.0685.3——
Cosmos 3 NanoModel Size=8B2026.0685.1——
BAGALType=Und., # Params=14B2025.0385——
Qwen3-VLVision Encoder=Qwen3ViT, LLM=Qwen3-8B, Model Adaptation Type=Open Source2026.0284——
PDropModels=BAGEL-7B, Avg. Tokens=61%2026.0584—94.1
SigLIP2 | Qwen-3-8BVision Encoder=SigLIP2, LLM=Qwen-3-8B, Model Adaptation Type=Internal E-commerce Adaptation2026.0282.5——
Cosmos-Reason2 8BModel Size=8B2026.0682.2——
Qwen2.5-VLVision Encoder=Qwen2.5ViT, LLM=Qwen2-7B, Model Adaptation Type=Open Source2026.0281.9——
RynnBrain 2BModel Size=2B2026.0681.9——
MiMo-Embodied 7BModel Size=7B2026.0681.4——
SigLIP2 | Qwen-3-4BVision Encoder=SigLIP2, LLM=Qwen-3-4B, Model Adaptation Type=Internal E-commerce Adaptation2026.0281——
VanillaModels=InternVL-U, Avg. Tokens=100%2026.0580.5—100
DualToken-7BType=Uni., # Params=7B, Resolution=384px2025.0380——
LLaVA-NeXTType=Und., # Params=34B2025.0379.3——
Show-o2Type=Uni., # Params=7B, Resolution=432px2025.0379.3——
G2TRModels=InternVL-U, Avg. Tokens=50%2026.0578.8—94.9
W-FastVModels=InternVL-U, Avg. Tokens=54%2026.0578.6—92.3
IVC-PruneModels=InternVL-U, Avg. Tokens=50%2026.0578.4—93.7
SigLIP | Gemma3-4BVision Encoder=SigLIP, LLM=Gemma3-4B, Model Adaptation Type=Internal E-commerce Adaptation2026.0278.3——
VSCANModels=InternVL-U, Avg. Tokens=50%2026.0578.1—92.2
FastVModels=InternVL-U, Avg. Tokens=54%2026.0577.9—92.9
PDropModels=InternVL-U, Avg. Tokens=61%2026.0577.7—92
SigLIP2 | Lilium-8BVision Encoder=SigLIP2, LLM=Lilium-8B, Model Adaptation Type=Internal E-commerce Adaptation2026.0277.4——
Qwen3-VL 2BModel Size=2B2026.0677.2——
SigLIP2 | e-Llama3.1-8BVision Encoder=SigLIP2, LLM=e-Llama3.1-8B, Model Adaptation Type=Internal E-commerce Adaptation2026.0276.9——
Cosmos 3 EdgeModel Size=2B2026.0676.6——
LLaVA-OVVision Encoder=SigLIP, LLM=Qwen2-7B, Model Adaptation Type=Open Source2026.0276.4——
DualToken-3BType=Uni., # Params=3B, Resolution=384px2025.0376.2——
SigLIP2 | Llama-3.1-8BVision Encoder=SigLIP2, LLM=Llama-3.1-8B, Model Adaptation Type=Internal E-commerce Adaptation2026.0275.8——
SigLIP2 | Lilium-4BVision Encoder=SigLIP2, LLM=Lilium-4B, Model Adaptation Type=Internal E-commerce Adaptation2026.0275.5——
DualToken-7BType=Uni., # Params=7B, Resolution=256px2025.0374.9——
Cosmos-Reason2 2BModel Size=2B2026.0673.6——
MUSE-VLType=Uni., # Params=7B, Resolution=256px2025.0372.1——
Qwen2.5ViT | e-Llama3.1-8BVision Encoder=Qwen2.5ViT, LLM=e-Llama3.1-8B, Model Adaptation Type=Internal E-commerce Adaptation2026.0271.7——
UniTokenType=Uni., # Params=7B, Resolution=384px2025.0371.1——
DualToken-3BType=Uni., # Params=3B, Resolution=256px2025.0370.9——
InternVL3.5-1B# Total Params=1.1B2025.1170.79——
JanusType=Uni., # Params=1.5B2025.0369.4——
LFM2-VL-1.6B# Total Params=1.6B2025.1169.16——
VILAType=Und., # Params=7B2025.0368.9——
TokenFlowType=Uni., # Params=13B, Resolution=384px2025.0368.9——
ShareGPT4VType=Und., # Params=7B2025.0368.8——
Gemma3Vision Encoder=SigLIP, LLM=Gemma3-4B, Model Adaptation Type=Open Source2026.0267.9——
LLaVA-NeXTType=Und., # Params=7B2025.0367.4——
Show-o2Type=Uni., # Params=1.5B, Resolution=432px2025.0367.4——
Gemma-4 E4BModel Size=E4B2026.0666.2——
SigLIP2 | Lilium-1BVision Encoder=SigLIP2, LLM=Lilium-1B, Model Adaptation Type=Internal E-commerce Adaptation2026.0264.7——
LLaVA-1.5Type=Und., # Params=7B2025.0364.3——
UniTokType=Uni., # Params=7B, Resolution=256px2025.0361.1——
Gemma-4 E2BModel Size=E2B2026.0660.2——
LLaVA-PhiType=Und., # Params=2.7B2025.0359.8——
Emu3Type=Uni., # Params=8B2025.0358.5——
LFM2-VL-450M# Total Params=0.45B2025.1155.76——
SmolVLM2-500M# Total Params=0.5B2025.1151.29——
InstructBLIPType=Und., # Params=7B2025.0336——
ChameleonType=Uni., # Params=7B2025.0331.1——
InfiMM-HD2024.030.716——
LLaVA 1.52024.030.677——
Sphinx-2K2024.030.659——
Qwen-VL-Chat2024.030.606——
Shikra2024.030.588——
Bunny-4BSize=Size < 4B2024.02—74.2—
Bunny-8BSize=4B < Size < 8B2024.02—77.2—
DeepSeek-VL-7BSize=4B < Size < 8B2024.02—73.2—
DivPruneBackbone=LLaVA-1.5-7B, Retained Tokens=1922025.09—62.54—
DivPruneBackbone=LLaVA-1.5-7B, Retained Tokens=1282025.09—62.03—
DivPruneBackbone=LLaVA-1.5-7B, Retained Tokens=642025.09—59.28—
FastVBackbone=LLaVA-1.5-7B, Retained Tokens=1922025.09—61.2—
FastVBackbone=LLaVA-1.5-7B, Retained Tokens=1282025.09—56.1—
FastVBackbone=LLaVA-1.5-7B, Retained Tokens=642025.09—48—
GPT-4V-1106Size=> 8B2024.02—75.1—
IDEFICS-9BLLM Param=8B, Type=Und. Only2024.11—48.2—
Idefics2Size=4B < Size < 8B2024.02—76.7—
Imp-v1.5-4B-Phi3Size=Size < 4B2024.02—73.3—
InstructBLIPLLM Param=7B, Type=Und. Only2024.11—36—
InstructBLIP-Vicuna-7BSize=4B < Size < 8B2024.02—36—
InternVL3.5-2B# Total Params=2.3B, Evaluation Toolkit=VLMEvalKit2025.11—78.18—
JanusLLM Param=1.3B, Type=Unified2024.11—69.4—
JanusFlowLLM Param=1.3B, Type=Unified2024.11—74.9—
LFM2-VL-3B# Total Params=3.0B, Evaluation Toolkit=VLMEvalKit2025.11—79.81—
LLAVALLM Param=7B, Type=Und. Only2024.11—38.7—
LLaVA-1.5-7BBackbone=LLaVA-1.5-7B, Retained Tokens=5762025.09—64.7—
LLaVA-NeXT-13BSize=> 8B2024.02—70—
LLaVA-NeXT-7BSize=4B < Size < 8B2024.02—67.4—
LLaVA-NeXT-Llama-3-8BSize=4B < Size < 8B2024.02—72.1—
LLAVA-PhiLLM Param=2.7B, Type=Und. Only2024.11—59.8—
LLaVA-v1.5LLM Param=7B, Type=Und. Only2024.11—64.3—
LLaVA-v1.5-13B (LORA)Size=> 8B2024.02—68.5—