List your company
ResearchFormerly WizWand

Multi-modal Understanding on MMVet (Accuracy)

85.67AccuracyVanilla MoE
26.837242.111157.38572.6589Aug 6, 2024Nov 17, 2024Mar 1, 2025Jun 12, 2025Sep 24, 2025Jan 5, 2026Apr 19, 2026
Updated 5mo ago

Evaluation results

MethodLinks
Vanilla MoEBackbone=QWEN3-VL-30B-A3B-INSTRUCT2026.0485.67
MACS (Ours)Backbone=QWEN3-VL-30B-A3B-INSTRUCT2026.0485.48
Vanilla MoEBackbone=INTERNVL3.5-30B-A3B2026.0485.43
MACS (Ours)Backbone=INTERNVL3.5-30B-A3B2026.0485.17
MACS (w/o Expanded)Backbone=QWEN3-VL-30B-A3B-INSTRUCT2026.0485.03
MACS (w/o Expanded)Backbone=INTERNVL3.5-30B-A3B2026.0484.66
CAI-MoE (Expanded)Backbone=QWEN3-VL-30B-A3B-INSTRUCT2026.0481.63
CAI-MoE (Expanded)Backbone=INTERNVL3.5-30B-A3B2026.0480.14
CAI-MoE (Token Drop)Backbone=QWEN3-VL-30B-A3B-INSTRUCT2026.0479.21
Vanilla MoEBackbone=KIMI-VL-A3B-INSTRUCT2026.0477.84
MACS (Ours)Backbone=KIMI-VL-A3B-INSTRUCT2026.0477.63
CAI-MoE (Token Drop)Backbone=INTERNVL3.5-30B-A3B2026.0477.19
MACS (w/o Expanded)Backbone=KIMI-VL-A3B-INSTRUCT2026.0477.12
GPT-4o2024.0876.2
CAI-MoE (Expanded)Backbone=KIMI-VL-A3B-INSTRUCT2026.0473.84
VAPOModel Scale=7B2025.0971.9
CAI-MoE (Token Drop)Backbone=KIMI-VL-A3B-INSTRUCT2026.0471.27
V-R1Model Scale=7B2025.0971.1
VLAAModel Scale=7B2025.0970
Top-k RoutingExpert skipping ratio (ρ)=0, Number of experts (k)=8, Backbone=InternVL-3.5-30B-A3B-HF2025.1169.93
Top-k RoutingExpert skipping ratio (ρ)=0.6, Number of experts (k)=3, Backbone=InternVL-3.5-30B-A3B-HF2025.1168.67
MC-MoEExpert skipping ratio (ρ)=0.6, Number of experts (k)=3, Backbone=InternVL-3.5-30B-A3B-HF2025.1168.67
MoDESExpert skipping ratio (ρ)=0.6, Number of experts (k)=3, Backbone=InternVL-3.5-30B-A3B-HF2025.1168.41
DiEPExpert skipping ratio (ρ)=0.6, Number of experts (k)=3, Backbone=InternVL-3.5-30B-A3B-HF2025.1168.13
MoDESExpert skipping ratio (ρ)=0.73, Number of experts (k)=2, Backbone=InternVL-3.5-30B-A3B-HF2025.1167.92
MoDESExpert skipping ratio (ρ)=0.85, Number of experts (k)=1, Backbone=InternVL-3.5-30B-A3B-HF2025.1167.8
NAEEExpert skipping ratio (ρ)=0.6, Number of experts (k)=3, Backbone=InternVL-3.5-30B-A3B-HF2025.1167.57
MC-MoEExpert skipping ratio (ρ)=0.73, Number of experts (k)=2, Backbone=InternVL-3.5-30B-A3B-HF2025.1166.41
MC-MoEExpert skipping ratio (ρ)=0.85, Number of experts (k)=1, Backbone=InternVL-3.5-30B-A3B-HF2025.1165.41
DiEPExpert skipping ratio (ρ)=0.85, Number of experts (k)=1, Backbone=InternVL-3.5-30B-A3B-HF2025.1165.32
DiEPExpert skipping ratio (ρ)=0.73, Number of experts (k)=2, Backbone=InternVL-3.5-30B-A3B-HF2025.1165.27
R1-OVModel Scale=7B2025.0965.2
Qwen-VLModel Scale=7B2025.0964.8
LLaVA-OneVision-72BModel Scale=72B2024.0863.7
Ovis2.5Parameter Scale=9B2026.0162.2
Qwen2-VL-7BModel=Qwen2-VL-7B2026.0162
VisionTrim on Qwen2-VL-7BModel=Qwen2-VL-7B, Tokens per image=approx. 1/3 original2026.0161.8
NAEEExpert skipping ratio (ρ)=0.73, Number of experts (k)=2, Backbone=InternVL-3.5-30B-A3B-HF2025.1161.38
Qwen3-VLParameter Scale=8B2026.0161.2
GLM-4.6V-FlashX2026.0161
Qwen2.5-VLParameter Scale=72B2026.0159.2
VisionTrim on LLaVA-OneVision-7BModel=LLaVA-OneVision-7B, Tokens per image=2432026.0159
LLaVA-OneVision-7BModel Scale=7B2024.0857.5
LLaVA-OneVision-7BModel=LLaVA-OneVision-7B, Tokens per image=7292026.0157.5
Top-k RoutingExpert skipping ratio (ρ)=0.73, Number of experts (k)=2, Backbone=InternVL-3.5-30B-A3B-HF2025.1156.42
InternVL3.5Parameter Scale=8B2026.0150.6
LATTETraining Data Size=293K2024.1250
GPT-4VVersion=V-Preview2024.0849.9
NAEEExpert skipping ratio (ρ)=0.85, Number of experts (k)=1, Backbone=InternVL-3.5-30B-A3B-HF2025.1149.52
LATTETraining Data Size=98K2024.1247.9
CogCoMTraining Data Size=70K2024.1246.1
Ostrakon-VLParameter Scale=8B2026.0136.4
Top-k RoutingExpert skipping ratio (ρ)=0.85, Number of experts (k)=1, Backbone=InternVL-3.5-30B-A3B-HF2025.1135.05
LLaVA-PlusTraining Data Size=158K2024.1235
LLaVA-OneVision-0.5BModel Scale=0.5B2024.0829.1