ImageNet-1K.

Current SOTA on ImageNet-1K. 2.1B parameters. Contrastive Captioner architecture.

Giant ViT variant. 1.8B parameters.

Giant ViT variant. 1.8B parameters.

SoViT-400m/14, shape-optimized ViT with 400M params. Finetuned on ImageNet-1K at 224px. 90.3% top-1 on IN-1K val. Surpasses ViT-g/14 (90.0%) at less than half the inference cost. NeurIPS 2023, paper revised Jan 2024. Source: arxiv:2305.13035 abstract.

EVA-02 ViT-L/14+ 304M params. MIM pre-training on Merged-38M, finetuned on IN-22K then IN-1K at 448x448. Source: timm results CSV (eva02_large_patch14_448.mim_m38m_ft_in22k_in1k). Paper: arxiv:2303.11331.

EVA ViT-Giant/14, 1B params. MIM pre-training on Merged-30M, finetuned on IN-22K then IN-1K at 560x560. Source: timm results CSV (eva_giant_patch14_560.m30m_ft_in22k_in1k). Paper: arxiv:2211.07636.

InternImage-H 1.08B params with deformable convolutions. IN-22K pretraining + joint ImageNet training, 640x640. Source: OpenGVLab/InternImage classification README. Paper: arxiv:2211.05778.

AIMv2-3B, multimodal autoregressive pre-training, 2.7B params, 448px. 89.5% top-1 on IN-1K val using attentive probing (frozen backbone + 2-layer attentive head). Apple, Nov 2024. Source: github.com/apple/ml-aim README table. Paper: arxiv:2411.14402.

Shape-Optimized SigLIP 400M, patch14, res 378. Contrastive pre-training on WebLI, finetuned on IN-1K. Source: timm results CSV (vit_so400m_patch14_siglip_378.webli_ft_in1k). Paper: arxiv:2303.15343.

Best pure ConvNet. 650M parameters. Trained with FCMAE.

Best pure ConvNet. 650M parameters. Trained with FCMAE.

ViT-H/14 CLIP pre-trained on LAION-2B, finetuned on IN-12K then IN-1K at 336px. Source: timm results CSV (vit_huge_patch14_clip_336.laion2b_ft_in12k_in1k). Paper: arxiv:2212.07143.

ConvNeXt-XXLarge, CLIP pre-trained on LAION-2B, soup finetuned on IN-1K. Source: timm results CSV (convnext_xxlarge.clip_laion2b_soup_ft_in1k). Paper: arxiv:2212.07143 (OpenCLIP).

Huge ViT variant. 632M parameters.

Huge ViT variant. 632M parameters.

InternViT-6B, 6B-param vision encoder, patch14, 224px. 88.23% Acc@1 on IN-1K val (50k images) via linear probing (frozen backbone + linear head). OpenGVLab, CVPR 2024 Oral. Source: Hugging Face model card OpenGVLab/InternViT-6B-448px-V2_5. Paper: arxiv:2312.14238.

Hierarchical Vision Transformer with shifted windows.

Hierarchical Vision Transformer with shifted windows.

Pretrained on ImageNet-21K, fine-tuned on 1K.

Pretrained on ImageNet-21K, fine-tuned on 1K.

MambaVision-L2, hybrid Mamba-Transformer backbone, 241M params. Finetuned on ImageNet-1K at 224px. 85.3% top-1. Sets new SOTA Pareto front for accuracy vs. throughput. NVIDIA, CVPR 2025. Source: arxiv:2407.08083 Table 1.

Data-efficient ViT with distillation. Trained on ImageNet-1K only.

Data-efficient ViT with distillation. Trained on ImageNet-1K only.

8.4x smaller than GPipe. 66M parameters.

Without distillation. Trained from scratch on ImageNet-1K.

28M parameters. Efficient variant.

28M parameters. Efficient variant.

Large ViT with ImageNet-21K pretraining.

Large ViT with ImageNet-21K pretraining.

Base ViT with ImageNet-21K pretraining.

Base ViT with ImageNet-21K pretraining.

ResNet Strikes Back. Modern training recipe on classic architecture.

ResNet Strikes Back. Modern training recipe on classic architecture.

10-crop evaluation. Original deep residual network.

Only 5.3M parameters. Baseline for compound scaling.

Standard torchvision baseline. 25M parameters.