CLIP
OpenAICLIP is OpenAI's 2021 contrastive image-text model and the origin of the whole cross-modal embedding line. An image encoder and a text encoder are trained to put matching pairs close together in one shared space, which makes image-text retrieval and zero-shot classification the same operation. Five years on it remains the single most downloaded vision model on the Hub, mostly as the frozen text or image tower inside diffusion and VLM stacks rather than as a retriever in its own right.
None of the four openai/clip-* Hub repos declares a license field; the MIT grant comes from the openai/CLIP repository LICENSE that distributes the checkpoints. The Xenova and timm re-uploads of these weights are mirrors, not products, and are excluded from the download sum.
Openness
3 high confidence- weights
- open(four ungated openai/clip-vit-* repos on the Hub, plus the original checkpoints served by the CLIP package)
- data
- closed(WIT-400M
- code
- partial(the openai/CLIP repository ships the model definition, inference and zero-shot evaluation notebooks, with no training pipeline)
- license
- MIT(OSI, from the openai/CLIP repository LICENSE
The category's centre of gravity, and the reason OpenCLIP exists. Weights under an OSI licence that anyone can run, a corpus the publisher says explicitly it will never release, and a repository that lets you use the model but not rebuild it.
- https://huggingface.co/openai/clip-vit-base-patch32/raw/main/README.md recorded 2026-10-06
'The model was trained on publicly available image-caption data ... A large portion of the data comes from our crawling of the internet'; 'We do not intend for this dataset to be used as the basis for any commercial or deployed model and will not be releasing the dataset.' Front matter carries `tags: vision` and no license field.
- https://raw.githubusercontent.com/openai/CLIP/main/LICENSE recorded 2026-10-06
'MIT License / Copyright (c) 2021 OpenAI'.
- https://raw.githubusercontent.com/openai/CLIP/main/README.md recorded 2026-10-06
usage is `clip.load` plus encode_image / encode_text and a zero-shot prediction example; the repository documents no training entry point.
Adoption
5 high confidenceAdoption is measured as Hugging Face downloads summed across CLIP's four openai/clip-vit-* repos, most of them for the base-patch32 checkpoint. Community re-uploads and format mirrors are excluded from the count.
- https://huggingface.co/api/models?author=openai&search=clip&limit=100&sort=downloads&direction=-1 recorded 2026-09-11
four openai clip repos with `downloads` (trailing 30 days) summing to 32,961,856.
Capability
2 high confidenceCLIP is English-only with a 77-token text context. On the zero-shot ImageNet benchmark it introduced, both OpenCLIP, the open reimplementation trained on open data, and SigLIP 2 now score higher, which puts CLIP a step below OpenCLIP despite its continued usefulness in narrower deployments.
- https://raw.githubusercontent.com/mlfoundations/open_clip/main/README.md recorded 2026-09-11
model table: 'ViT-L-14-quickgelu (Original CLIP) | WIT | 224px | 13B | 75.5%' ImageNet zero-shot acc., against ViT-H-14 LAION-2B at 78.0% and ViT-bigG-14 at 80.1%.
- https://raw.githubusercontent.com/openai/CLIP/main/README.md recorded 2026-09-11
'We found CLIP matches the performance of the original ResNet50 on ImageNet "zero-shot" without using any of the original 1.28M labeled examples'.
Verified 2026-09-11