SAVRN's Take
Every model here does one job: it takes an image in and hands back a set of numbers that describe it, so downstream systems can search or compare by those numbers. Six are indexed, all Vision Transformers, and they are tiny next to anything else we price. The 22M-parameter dinov2-small needs 0.1 GB at 16-bit; dinov2-large, at 304M parameters, needs 0.7 GB, and that is the whole range. None fills a card, so the question is how many image streams one GPU carries, not how many GPUs you need. The cheapest hosted entry for all six is a single MI300X at $1.85 an hour.
AI at Meta wrote three of the six and holds the top two download slots: dinov2-base at 87M parameters draws 3,364,437 downloads a month and dinov2-small draws 3,240,723. Google's vit-base-patch16-224-in21k, pre-trained on ImageNet-21k's 14 million images across 21,843 classes at 224 by 224, sits third at 1,800,290. Five of the six carry Apache 2.0, so commercial use, modification and redistribution need no further permission. The exception is Camenduru's 303M-parameter dinov3-vitl16-pretrain-lvd1689m, filed under a license the index records only as other, and a buyer reads those terms line by line before production.
Two more checks. Match the serving stack: five load through transformers, the timm build through timm. And skip quantization; dinov2-large drops from 0.7 GB to 0.4 GB at 8-bit and 0.2 GB at 4-bit, the small ones round to zero. Run at 16-bit, put the effort into batch size, and price the job in images per hour on that $1.85 card.
SAVRN Research, 2026-09-18
