List your company
ResearchFormerly WizWand

Accuracy on MMMU

83.4Accuracy (MMMU)S1-VL-32B-RL
19.12835.81452.569.186Oct 13, 2025Nov 25, 2025Jan 7, 2026Feb 19, 2026Apr 3, 2026May 16, 2026Jun 29, 2026
Updated 4mo ago

Evaluation results

MethodLinks
S1-VL-32B-RL2026.0683.4
S1-Omni-Image2026.0682.67
S1-VL-32B-SFT2026.0682.5
GPT-52026.0681.22
Qwen3-VL-235B-A22B-Thinking2026.0677.89
Qwen3-VL-32B-Thinking2026.0676
BF16Precision=BF162026.0660.11
Qwen3-VL-32B-InstructBackbone model=Qwen3-VL-32B-Instruct2026.0558
MODEQuantization=W32026.0657.33
IVR-R1Backbone model=Qwen3-VL-4B2026.0556
VEQ-MAQuantization=W32026.0655.49
MC-MoEQuantization=W32026.0654.56
Vision-R1Backbone model=Qwen3-VL-4B2026.0554.1
IVR-R1Backbone model=Qwen2.5-VL-7B2026.0553.6
GPTQQuantization=W32026.0652.56
Vision-SR1Backbone model=Qwen3-VL-4B2026.0552.4
QWEN2.5-VL-7B-INSTRUCT (PALETTE)Allowed Domain=PH2026.0552
QWEN2.5-VL-7B-INSTRUCT (PALETTE)Allowed Domain=PV2026.0551.7
QWEN2.5-VL-7B-INSTRUCT (PALETTE)Allowed Domain=HS2026.0551.3
QWEN2.5-VL-7B-INSTRUCT (PALETTE)Allowed Domain=Fraud2026.0551.1
Vision-R1Backbone model=Qwen2.5-VL-7B2026.0550.9
QWEN2.5-VL-7B-INSTRUCTAllowed Domain=Base2026.0550.9
QWEN2.5-VL-7B-INSTRUCT (PALETTE)Allowed Domain=Sex2026.0550.9
QWEN2.5-VL-7B-INSTRUCT (PALETTE)Allowed Domain=IA2026.0550.8
SplitQBackbone=Qwen2.5-VL-7B, Bits=W4A82026.0549.1
MODEQuantization=W22026.0648.67
Thyme-VL2026.0648.66
Supervised Fine-tuning (before RL)Backbone model=Qwen3-VL-4B2026.0548.4
Vision-SR1Backbone model=Qwen2.5-VL-7B2026.0547.6
Zero-shot Inference (before RL)Backbone model=Qwen3-VL-4B2026.0547.4
SplitQBackbone=Qwen2.5-VL-7B, Bits=W4A42026.0546.9
MASQBackbone=Qwen2.5-VL-3B, Bits=W4A82026.0546.7
FP16Backbone=Qwen2.5-VL-7B, Bits=W16A162026.0546.7
SplitQBackbone=Qwen2.5-VL-3B, Bits=W4A82026.0546.3
VEQ-MAQuantization=W22026.0646.11
Supervised Fine-tuning (before RL)Backbone model=Qwen2.5-VL-7B2026.0546.1
MC-MoEQuantization=W22026.0644.89
SplitQBackbone=Qwen2.5-VL-7B, Bits=W3A32026.0543.9
SplitQBackbone=Qwen2.5-VL-3B, Bits=W4A42026.0543.7
MASQBackbone=Qwen2.5-VL-7B, Bits=W4A82026.0543.4
MBQBackbone=Qwen2.5-VL-7B, Bits=W4A82026.0543.3
FP16Backbone=Qwen2.5-VL-3B, Bits=W16A162026.0542.2
MBQBackbone=Qwen2.5-VL-3B, Bits=W4A82026.0541.2
SplitQBackbone=Qwen2.5-VL-3B, Bits=W3A32026.0539.8
LLaVA-NEXT-3BBackbone=LLaVA-NEXT-3B2026.0539.6
LBRBackbone=LLaVA-NEXT-3B2026.0538.8
LLaVA-1.5 + FG-CLIP 2Vision Encoder=FG-CLIP 2, LMM Base Architecture=LLaVA-1.52025.1038.1
LBRBackbone=LLaVA-1.5-13B2026.0537.9
SQBackbone=Qwen2.5-VL-7B, Bits=W4A82026.0537.8
LBRBackbone=LLaVA-1.5-7B2026.0537.2
LLaVA-1.5 + SigLIP 2Vision Encoder=SigLIP 2, LMM Base Architecture=LLaVA-1.52025.1037
LLaVA-1.5 + Meta CLIP 2Vision Encoder=Meta CLIP 2, LMM Base Architecture=LLaVA-1.52025.1037
LLaVA-1.5-13BBackbone=LLaVA-1.5-13B2026.0536.7
LLaVA-1.5-7BBackbone=LLaVA-1.5-7B2026.0535.7
LLaVA-1.5 + CLIPVision Encoder=CLIP, LMM Base Architecture=LLaVA-1.52025.1035.7
SplitQBackbone=Qwen2.5-VL-7B, Bits=W3A22026.0535.6
Qwen2.5-VL-72B-InstructBackbone model=Qwen2.5-VL-72B-Instruct2026.0534.9
SplitQBackbone=Qwen2.5-VL-3B, Bits=W3A22026.0533.1
OPPOTraining=Oriented Pickup Preference Optimization, Base Model=LLaVA-NeXT2026.0632.3
LLaVA-NeXT (Baseline)Training=Vanilla2026.0631.8
mDPOTraining=Preference Optimization, Base Model=LLaVA-NeXT2026.0631.5
CHiPTraining=Preference Optimization, Base Model=LLaVA-NeXT2026.0631
DPOTraining=Preference Optimization, Base Model=LLaVA-NeXT2026.0630.9
MASQBackbone=Qwen2.5-VL-3B, Bits=W4A42026.0526.7
MBQBackbone=Qwen2.5-VL-7B, Bits=W4A42026.0526.7
SQBackbone=Qwen2.5-VL-3B, Bits=W4A82026.0525.6
MBQBackbone=Qwen2.5-VL-3B, Bits=W4A42026.0525
MASQBackbone=Qwen2.5-VL-7B, Bits=W4A42026.0525
SQBackbone=Qwen2.5-VL-7B, Bits=W4A42026.0524.8
SQBackbone=Qwen2.5-VL-3B, Bits=W4A42026.0523.3
Qwen2.5-VL-32B-InstructBackbone model=Qwen2.5-VL-32B-Instruct2026.0522.9
GPTQQuantization=W22026.0622.78
Zero-shot Inference (before RL)Backbone model=Qwen2.5-VL-7B2026.0521.6