Image Classification on ImageNet (Top-1 Acc)
90.94Top-1 AccuracyModel Soups ViT-G
Evaluation Results
| Method | Links | |
|---|---|---|
| Model Soups ViT-GBackbone=ViT-G2025.07 | 90.94 | |
| PE-coreBackbone=G/14, Evaluation Protocol=lin2026.04 | 89.5 | |
| DINOv2Backbone=g/14, Evaluation Protocol=lin2026.04 | 87.3 | |
| PE-coreBackbone=G/14, Evaluation Protocol=KNN2026.04 | 86.8 | |
| TIPSv2Backbone=g/14, Evaluation Protocol=lin2026.04 | 86.8 | |
| TIPSBackbone=g/14, Evaluation Protocol=lin2026.04 | 86.2 | |
| FRANCABackbone=G/14, Evaluation Protocol=lin2026.04 | 85.9 | |
| PE-coreModel=PE-core G/14, Zero-shot protocol=true2026.04 | 85.4 | |
| SigLIP2Model=SigLIP2 g/16, Zero-shot protocol=true2026.04 | 85 | |
| OpenCLIP HScale=H2025.07 | 84.37 | |
| CLIPBackbone=L/14, Evaluation Protocol=lin2026.04 | 84.3 | |
| DualOptModel=Swin-B2026.04 | 83.89 | |
| TIPSv2Backbone=g/14, Evaluation Protocol=KNN2026.04 | 83.7 | |
| SILCModel=SILC G/16, Zero-shot protocol=true2026.04 | 83.7 | |
| FastViT-SA36Latency (ms)=44.4, Power Consumption (W)=9.02026.03 | 83.6 | |
| LowFormer-B3Latency (ms)=32.5, Power Consumption (W)=9.22026.03 | 83.6 | |
| DINOv2Backbone=g/14, Evaluation Protocol=KNN2026.04 | 83.6 | |
| TIPSBackbone=g/14, Evaluation Protocol=KNN2026.04 | 83.3 | |
| OriginModel=Swin-B2026.04 | 83.05 | |
| LowFormer-E3Latency (ms)=25.0, Power Consumption (W)=9.12026.03 | 83 | |
| FRANCABackbone=G/14, Evaluation Protocol=KNN2026.04 | 83 | |
| FastViT-SA24Latency (ms)=30.9, Power Consumption (W)=9.02026.03 | 82.6 | |
| PreNAS-B*#Params=54.0M, FLOPs=11.0G2026.05 | 82.6 | |
| TAS-LoRA-B#Params=54.0M, FLOPs=11.0G2026.05 | 82.6 | |
| DualOptModel=ViT-B2026.04 | 82.58 | |
| ConViT-B#Params=86.0M, FLOPs=17.0G2026.05 | 82.4 | |
| AutoFormer-B#Params=54.0M, FLOPs=11.0G2026.05 | 82.4 | |
| AutoFormer-B†#Params=54.0M, FLOPs=11.0G2026.05 | 82.4 | |
| GLiT-B*#Params=96.0M, FLOPs=17.0G2026.05 | 82.3 | |
| TF-TAS-B*#Params=54.0M, FLOPs=12.0G2026.05 | 82.2 | |
| TF-TAS-S*#Params=22.8M, FLOPs=5.0G2026.05 | 81.9 | |
| TAS-LoRA-S#Params=22.9M, FLOPs=5.0G2026.05 | 81.9 | |
| FocusFormer-B*#Params=52.7M, FLOPs=11.0G2026.05 | 81.9 | |
| PreNAS-S*#Params=22.9M, FLOPs=5.1G2026.05 | 81.8 | |
| DeiT-B#Params=86.0M, FLOPs=18.0G2026.05 | 81.8 | |
| DeiT-BaseParams (M)=86.5, FLOPs (G)=17.56, Structured Compression=false2025.05 | 81.8 | |
| FPQuantization=W4A4, Backbone=DeiT-B2026.05 | 81.8 | |
| FPQuantization=W3A3, Backbone=DeiT-B2026.05 | 81.8 | |
| FPQuantization=W2A4, Backbone=DeiT-B2026.05 | 81.8 | |
| AutoFormer-S#Params=22.9M, FLOPs=5.1G2026.05 | 81.7 | |
| PVT-Large#Params=61.0M, FLOPs=9.8G2026.05 | 81.7 | |
| LowFormer-E2Latency (ms)=14.7, Power Consumption (W)=9.32026.03 | 81.6 | |
| FocusFormer-S*#Params=23.7M, FLOPs=5.0G2026.05 | 81.6 | |
| AutoFormer-S†#Params=22.9M, FLOPs=5.1G2026.05 | 81.6 | |
| TNT-S#Params=23.8M, FLOPs=5.2G2026.05 | 81.5 | |
| DINOv2 (dino.txt)Model=DINOv2 (dino.txt) L/14, Zero-shot protocol=true2026.04 | 81.4 | |
| OriginModel=ViT-B2026.04 | 81.34 | |
| FP16Backbone=DeiT-Base, Quantization=FP162026.04 | 81.3 | |
| ConViT-S#Params=27.0M, FLOPs=5.4G2026.05 | 81.3 | |
| LowFormer-B1.5Latency (ms)=18.1, Power Consumption (W)=8.32026.03 | 81.2 | |
| EdgeViT-SLatency (ms)=24.6, Power Consumption (W)=8.02026.03 | 81 | |
| EfficientMod-sLatency (ms)=30.5, Power Consumption (W)=7.92026.03 | 81 | |
| TIPSv2Model=TIPSv2 g/14 (ours), Zero-shot protocol=true2026.04 | 80.7 | |
| FastViT-SA12Latency (ms)=17.5, Power Consumption (W)=8.82026.03 | 80.6 | |
| GLiT-S*#Params=24.6M, FLOPs=4.4G2026.05 | 80.5 | |
| ViTAS-F#Params=27.6M, FLOPs=6.0G2026.05 | 80.5 | |
| GPTAQ-MARRQuantization=W4A4, Backbone=DeiT-B2026.05 | 80.11 | |
| OpenCLIPModel=OpenCLIP G/14, Zero-shot protocol=true2026.04 | 80.1 | |
| LowFormer-B1Latency (ms)=11.7, Power Consumption (W)=7.72026.03 | 79.9 | |
| TIPSModel=TIPS g/14, Zero-shot protocol=true2026.04 | 79.9 | |
| DeiT-S#Params=22.1M, FLOPs=4.7G2026.05 | 79.9 | |
| FPQuantization=W4A4, Backbone=DeiT-S2026.05 | 79.85 | |
| FPQuantization=W3A3, Backbone=DeiT-S2026.05 | 79.85 | |
| FPQuantization=W2A4, Backbone=DeiT-S2026.05 | 79.85 | |
| FP16Backbone=DeiT-Small, Quantization=FP162026.04 | 79.8 | |
| CLIPBackbone=L/14, Evaluation Protocol=KNN2026.04 | 79.8 | |
| DeiT-SmallParams (M)=22.1, FLOPs (G)=4.60, Structured Compression=false2025.05 | 79.8 | |
| ViT-B#Params=86.0M, FLOPs=18.0G2026.05 | 79.7 | |
| FAR-BaseParams (M)=83.2, FLOPs (G)=17.31, Structured Compression=false2025.05 | 79.7 | |
| GPTAQQuantization=W4A4, Backbone=DeiT-B2026.05 | 79.62 | |
| MobileOne-S4Latency (ms)=18.6, Power Consumption (W)=9.32026.03 | 79.4 | |
| MAGE2026.03 | 78.9 | |
| LowFormer-E1Latency (ms)=6.2, Power Consumption (W)=9.52026.03 | 78.8 | |
| ViT-S#Params=22.1M, FLOPs=4.7G2026.05 | 78.8 | |
| GPTQQuantization=W4A4, Backbone=DeiT-B2026.05 | 78.79 | |
| ResComp-MARRQuantization=W4A4, Backbone=DeiT-B2026.05 | 78.68 | |
| PTQ4ViTQuantization=W4A4, Backbone=DeiT-B2026.05 | 78.53 | |
| RepQ-ViTQuantization=W4A4, Backbone=DeiT-B2026.05 | 78.37 | |
| EfficientMod-xsLatency (ms)=17.7, Power Consumption (W)=7.82026.03 | 78.3 | |
| ResCompQuantization=W4A4, Backbone=DeiT-B2026.05 | 78.27 | |
| A-SelecT2026.03 | 78.2 | |
| COSMICInference-time adaptation=true2026.06 | 78.2 | |
| MobileOne-S3Latency (ms)=11.8, Power Consumption (W)=9.32026.03 | 78.1 | |
| GPTAQBackbone=DeiT-Base, Quantization=W4A42026.04 | 78.1 | |
| LLQROptimizer=SGDM, Block=E-KFAC, Div.=NGD, Backbone=ResNet-50, Time=×1.0322026.05 | 78.05 | |
| CAEBackbone=DeiT-Base, Quantization=W4A42026.04 | 78 | |
| LLQROptimizer=SGDM, Block=K-FAC, Div.=NGD, Backbone=ResNet-50, Time=×1.0272026.05 | 77.98 | |
| LAGOBackbone=L/14, Zero-shot=true2026.05 | 77.97 | |
| BiFTABackbone=L/14, Zero-shot=true2026.05 | 77.82 | |
| FAR-SmallParams (M)=23.9, FLOPs (G)=4.82, Structured Compression=false2025.05 | 77.8 | |
| LLQROptimizer=SGDM, Block=K-FAC, Div.=Newton, Backbone=ResNet-50, Time=×1.0272026.05 | 77.78 | |
| GPTQBackbone=DeiT-Base, Quantization=W4A42026.04 | 77.6 | |
| EdgeViT-XSLatency (ms)=16.0, Power Consumption (W)=8.12026.03 | 77.5 | |
| FAR-Base (Pruned)Params (M)=73.0, FLOPs (G)=14.80, Structured Compression=true2025.05 | 77.5 | |
| SGDMOptimizer=SGDM, Backbone=ResNet-50, Time=×1.002026.05 | 77.42 | |
| WCABackbone=L/14, Zero-shot=true2026.05 | 77.32 | |
| SDXL2026.03 | 77.2 | |
| GPT-4o2025.07 | 77.2 | |
| PreNAS-Ti*#Params=5.9M, FLOPs=1.4G2026.05 | 77.1 | |
| DifFeed2026.03 | 77 |