Image Classification Models

SAVRN's Take

Capacity is not the constraint in this group. The 63 models under image classification run from 471,768 parameters, a timm test ResNet meant for sanity checks, up to 198M parameters for a ConvNeXt-V2 large. Even that top end needs 0.5 GB at 16-bit and 0.1 GB at 4-bit, and from the smallest entry to the largest the cheapest host we price is the same, one MI300X at $1.85 an hour. The hardware question is how many classifiers you stack on one card, not which card.

One publisher dominates. PyTorch Image Models accounts for 41 of the 63 entries, with Microsoft, AI at Meta and Rizvan Dwikifirdaus at 2 apiece, and holds the top two download slots: mobilenetv3_small_100.lamb_in1k at 17,686,947 a month with 3M parameters, then efficientnet_b3.ra2_in1k at 12,700,426 with 12M. Google's vit-base-patch16-224, 87M parameters, is the most downloaded transformer option at 7,074,109. Falcons.ai's nsfw_image_detection, a fine-tuned ViT of 86M, sits at 3,413,260, then timm's resnet18 and resnet50 at 2,003,638 and 1,929,644, and Rizvan Dwikifirdaus's gender-classification, another 86M ViT, at 1,422,012.

Licensing is where we would spend the diligence time. Apache-2.0 covers 44 of the 63 and MIT another 7, so most of the group arrives on permissive terms. The rest need a closer read before any commercial deployment: 6 carry CC-BY-NC-4.0, including that 198M ConvNeXt-V2 large, one is CC-BY-NC-ND-4.0, and 3 state no license, the gender-classification model among them. Check the license on the exact checkpoint, not the family name, and confirm whether it loads in timm or transformers, since the entries here split between the two.

SAVRN Research, 2026-09-18