SigLIP
Google'/%3e%3cdefs%3e%3cpattern%20id='pattern0_1240_293949'%20patternUnits='userSpaceOnUse'%20patternTransform='matrix(11.2%200%200%2011.2%203.93617%200)'%20preserveAspectRatio='none'%20viewBox='0%200%2016%2016'%20width='1'%20height='1'%3e%3cg%20id='pattern0_1240_293949_inner'%3e%3crect%20width='16'%20height='16'%20fill='%23f88376'/%3e%3crect%20x='6'%20width='2'%20height='2'%20fill='%23F7F6F6'/%3e%3crect%20x='6'%20y='8'%20width='2'%20height='2'%20fill='%23F7F6F6'/%3e%3crect%20x='4'%20y='2'%20width='2'%20height='2'%20fill='%23F7F6F6'/%3e%3crect%20x='4'%20y='10'%20width='2'%20height='2'%20fill='%23F7F6F6'/%3e%3crect%20x='2'%20y='4'%20width='2'%20height='2'%20fill='%23F7F6F6'/%3e%3crect%20x='2'%20y='12'%20width='2'%20height='2'%20fill='%23F7F6F6'/%3e%3crect%20y='6'%20width='2'%20height='2'%20fill='%23F7F6F6'/%3e%3crect%20y='14'%20width='2'%20height='2'%20fill='%23F7F6F6'/%3e%3crect%20x='14'%20width='2'%20height='2'%20fill='%23F7F6F6'/%3e%3crect%20x='14'%20y='8'%20width='2'%20height='2'%20fill='%23F7F6F6'/%3e%3crect%20x='12'%20y='2'%20width='2'%20height='2'%20fill='%23F7F6F6'/%3e%3crect%20x='12'%20y='10'%20width='2'%20height='2'%20fill='%23F7F6F6'/%3e%3crect%20x='10'%20y='4'%20width='2'%20height='2'%20fill='%23F7F6F6'/%3e%3crect%20x='10'%20y='12'%20width='2'%20height='2'%20fill='%23F7F6F6'/%3e%3crect%20x='8'%20y='6'%20width='2'%20height='2'%20fill='%23F7F6F6'/%3e%3crect%20x='8'%20y='14'%20width='2'%20height='2'%20fill='%23F7F6F6'/%3e%3c/g%3e%3c/pattern%3e%3c/defs%3e%3c/svg%3e)
SigLIP replaces CLIP's softmax contrastive loss with a pairwise sigmoid loss, which needs no global view of the batch and so trains well at both very large and quite small batch sizes. SigLIP 2 folds captioning pretraining, self-distillation, masked prediction and online data curation into the same recipe and adds multilingual coverage and native-aspect-ratio variants. The checkpoints serve two audiences at once: image-text retrieval in their own right, and the default frozen vision tower for a large share of current open VLMs.
The Hub cards for SigLIP 1 were written by Hugging Face rather than by the release team, and both generations put their evaluation tables in images; the numbers used here come from big_vision's own SigLIP 2 checkpoint table instead. Verified 2026-09-11 via the siglip2-base and siglip-so400m model cards, the big_vision README and its SigLIP 2 checkpoint README.
Openness
3 high confidence- weights
- open(44 ungated Apache-2.0 google/siglip* repos on the Hub, plus the .npz checkpoints served directly from the big_vision bucket)
- data
- closed(WebLI, a private Google web corpus
- code
- partial(big_vision ships a SigLIP trainer and a LiT/COCO demo config, but records 'code TODO' against the SigLIP paper entry and publishes no WebLI pretraining config)
- license
- Apache-2.0(OSI)
Apache-2.0 weights with no gate, which is more permissive than most of Google's model releases, over a corpus that is named and nothing more. big_vision gets closer to a pipeline than CLIP's repository does - there is a real SigLIP trainer in it - but the entry for the SigLIP paper says 'code TODO' and the config that would reproduce a WebLI run is not there, so this is partial rather than open.
- https://huggingface.co/google/siglip2-base-patch16-224/raw/main/README.md recorded 2026-09-11
`license: apache-2.0`; 'SigLIP 2 is pre-trained on the WebLI dataset (Chen et al., 2023)'; evaluation results supplied as an image.
- https://huggingface.co/google/siglip-so400m-patch14-384/raw/main/README.md recorded 2026-09-11
`license: apache-2.0`; 'SigLIP is pre-trained on the WebLI dataset'; 'first released in this repository' pointing at google-research/big_vision.
- https://raw.githubusercontent.com/google-research/big_vision/main/README.md recorded 2026-09-11
the SigLIP paper entry reads 'Resources: colab and models, code TODO'; an 'Image-text training with SigLIP' section launches big_vision.trainers.proj.image_text.siglip with the siglip_lit_coco.py config.
Adoption
5 high confidence11,451,605 downloads in the trailing 30 days across the 44 google/siglip* repos (siglip2-giant-opt-patch16-384 2,628,928; siglip-base-patch16-224 1,914,186; siglip2-base-patch16-224 1,546,725). The timm re-hosts of the same checkpoints are excluded as mirrors.
- https://huggingface.co/api/models?author=google&search=siglip&limit=100&sort=downloads&direction=-1 recorded 2026-09-11
44 google siglip repos with `downloads` (trailing 30 days) summing to 11,451,605.
Capability
4 high confidenceRung 4, competitive frontier: current generation, multilingual, five points of zero-shot ImageNet clear of the best OpenCLIP checkpoint, and the default vision tower for much of the open VLM stack. Short of 5 because the category reserves that for the acknowledged leader, and on the same big_vision instrument the gap to the top of the open image-text field is narrow rather than absent. Recorded `at` ColPali - a different modality but the other frontier-tier release in this cluster - because the text anchor is not a comparable measurement. Placed against the rung definition rather than a peer: no single measurement reports this product and a category peer together - zero-shot ImageNet against ColPali's ViDoRe - and rule (d) forbids an edge without one.
- https://raw.githubusercontent.com/google-research/big_vision/main/big_vision/configs/proj/image_text/README_siglip2.md recorded 2026-09-11
checkpoint table, INet 0-shot / COCO T-I / COCO I-T: g-opt/16 384 -> 85.0 / 56.1 / 72.8; So400m/16 512 -> 84.3 / 56.0 / 71.3; B/16 224 -> 78.2 / 52.1 / 68.9.
- https://huggingface.co/google/siglip2-base-patch16-224/raw/main/README.md recorded 2026-09-11
'SigLIP 2 models outperform their SigLIP counterparts at all model scales in core capabilities, including zero-shot classification, image-text retrieval, and transfer performance'.
Verified 2026-09-11