SAVRN
Search Contact SAVRN

SAVRN Model Hub · Comparisons

CLIP-ViT-B-32-laion2B-s34B-b79K vs siglip-base-patch16-224

CLIP-ViT-B-32-laion2B-s34B-b79K has 151M parameters and siglip-base-patch16-224 has 203M parameters; CLIP-ViT-B-32-laion2B-s34B-b79K is released under MIT License and siglip-base-patch16-224 under Apache License 2.0; at 16-bit, CLIP-ViT-B-32-laion2B-s34B-b79K needs about 0.4 GB (1x MI300X from $1.85 an hour) and siglip-base-patch16-224 about 0.5 GB (1x MI300X from $1.85 an hour).

Published metadata for 2 models, each read from its own repository.
Field CLIP-ViT-B-32-laion2B-s34B-b79K
laion/CLIP-ViT-B-32-laion2B-s34B-b79K
siglip-base-patch16-224
google/siglip-base-patch16-224
Publisher LAION eV Google
Task Zero shot image classification Zero shot image classification
Modality Other Other
Parameters, as reported 151M parameters 203M parameters
Architecture CLIPModel SiglipModel
Library open_clip transformers
Context length 77 tokens Not stated
Repository size 2.4 GB 1.6 GB
Artifact formats safetensors, pytorch safetensors, pytorch
License mit apache-2.0
Access Open weights, no gate Open weights, no gate
Memory at 16-bit (weights and margin) 0.4 GB 0.5 GB
Cheapest GPUs at 16-bit, per hour 1x MI300X, $1.85 1x MI300X, $1.85
Memory at 4-bit (weights and margin) 0.1 GB 0.1 GB
Cheapest GPUs at 4-bit, per hour 1x MI300X, $1.85 1x MI300X, $1.85
Revision viewed 1a25a446712b 7fd15f0689c7
Downloads reported by the hub 3.8M 2M
Last observed 2026-09-18 2026-09-18

An evaluation row appears only where at least two of these models report the same benchmark with the same stated configuration, metric, unit and setup. Different evaluators stay named in each cell. Values are shown as reported: no unit conversion, no ranking.

SAVRN's Notes on CLIP-ViT-B-32-laion2B-s34B-b79K

On disk it is 2.4 GB across 14 files, stored in float32; in 16-bit it needs 0.4 GB, and 8-bit halves that to 0.2 GB. The cheapest setup we list, one MI300X with 192 GB at $1.85 an hour, dwarfs a 151M-parameter zero-shot image classifier, so share the card. The text side reads 77 tokens, which keeps the labels you pair with images short, and the training data was the LAION-2B English subset of LAION-5B.

MIT is a short license: commercial use, modification and redistribution are allowed as long as the copyright and permission notices ride along. Set that against the publisher's framing, which calls the model a research output for research communities; the license allows a production deployment, the stated intent does not describe one. Check the serving stack too: the library is open_clip, not transformers, and the one paper on file is arXiv:1910.04867.

SAVRN's Notes on siglip-base-patch16-224

Look at your images at 224 by 224 first, the resolution of Google's WebLi pretraining. Labels go in as text and images get scored against them; no classifier trained. On disk, 1.6 GB in float32 for 203M parameters; loaded at 16-bit the footprint is 0.4 GB of weights inside 0.5 GB of memory, and the cheapest card the Index lists, one MI300X with 192 GB at $1.85 an hour, charges you for the hour, not the memory.

Google released it on September 30, 2023, but the description was written by someone other than the release team, so read the two papers instead, arXiv:2303.15343 on the sigmoid loss and arXiv:2209.06794 on PaLI. Access is open and Apache 2.0 permits commercial use, modification and redistribution, provided the notices and NOTICE file stay and significant changes are stated; nothing there blocks an in-house deployment.

Questions

Which is larger, CLIP-ViT-B-32-laion2B-s34B-b79K or siglip-base-patch16-224?

siglip-base-patch16-224 (203M parameters) is larger than CLIP-ViT-B-32-laion2B-s34B-b79K (151M parameters), by the parameter counts their publishers report.

Which is cheaper to run, CLIP-ViT-B-32-laion2B-s34B-b79K or siglip-base-patch16-224?

At 4-bit, CLIP-ViT-B-32-laion2B-s34B-b79K fits on 1x MI300X from $1.85 an hour and siglip-base-patch16-224 on 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.

Can I use CLIP-ViT-B-32-laion2B-s34B-b79K commercially?

Yes. CLIP-ViT-B-32-laion2B-s34B-b79K is released under MIT License. The MIT License is a short permissive license. It permits commercial use, modification and redistribution, provided the copyright notice and permission notice are included.

Can I use siglip-base-patch16-224 commercially?

Yes. siglip-base-patch16-224 is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.

Related Comparisons