OpenCLIP
LAIONOpenCLIP is the open reimplementation of CLIP: a full contrastive pretraining pipeline plus a family of image-text checkpoints trained on LAION's published corpora rather than on a private crawl. Because both halves are in the open, it is the line that made CLIP-style scaling reproducible - the ViT-bigG/14 checkpoint reaches 80.1% zero-shot ImageNet, above anything OpenAI released - and it is the reference implementation almost every later CLIP variant is trained and evaluated with.
Two organizations in one product: the open_clip training code is maintained under mlfoundations, the checkpoints and the LAION corpora are published by LAION. The corpus history is not clean - laion/laion2B-en no longer resolves under its own name and now redirects to the re-filtered laion/relaion2B-en-research-safe, which sits behind an auto-accept gate. Verified 2026-09-11 via the open_clip README and LICENSE, the ViT-H/14 model card and the Hugging Face API.
Openness
5 high confidence- weights
- open(56 ungated laion/CLIP-* checkpoints on the Hub - 50 MIT, 4 Apache-2.0 and 2 declaring none)
- data
- open(the DataComp-1B corpus behind the DataComp.XL checkpoints is published ungated under CC-BY-4.0
- code
- open(open_clip ships the full contrastive pretraining pipeline - data loading, training loop, distributed launch scripts and ImageNet zero-shot evaluation)
- license
- MIT(OSI)
The top rung, and the clearest case for it in this set: the corpus, the code that consumed it and an OSI licence over the weights. The one wrinkle is that the corpus these checkpoints were actually trained on was withdrawn in 2023 and the name now resolves to the re-filtered ReLAION release, which is auto-gated - a click-through, not an approval, so it is still released rather than merely documented.
- https://huggingface.co/api/datasets/mlfoundations/datacomp_1b recorded 2026-09-11
gated: false; license cc-by-4.0 - DataComp-1B, the corpus behind the DataComp.XL checkpoints, is published and ungated. This is the firmer footing for data:open than the LAION-2B path, whose original name now redirects to a re-filtered successor behind a gate.
- https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K/raw/main/README.md recorded 2026-09-11
`license: mit`; 'A CLIP ViT-H/14 model trained with the LAION-2B English subset of LAION-5B ... using OpenCLIP'; a Training Data section naming the 2-billion-sample English subset.
- https://huggingface.co/api/datasets/laion/relaion2B-en-research-safe recorded 2026-09-11
`id: laion/relaion2B-en-research-safe`, `gated: auto`, `private: false`, `disabled: false` - and the request for laion/laion2B-en redirects here.
- https://raw.githubusercontent.com/mlfoundations/open_clip/main/README.md recorded 2026-09-11
'This repository is focused on training CLIP models'; documents `python -m open_clip_train.main` with train-data, epochs and imagenet-val arguments, plus single-node and SLURM multi-node launch recipes.
- https://raw.githubusercontent.com/mlfoundations/open_clip/main/LICENSE recorded 2026-09-11
the MIT permission grant verbatim - 'to deal in the Software without restriction' - over a 2012-2021 copyright by the open_clip authors.
Adoption
5 medium confidence10,909,706 downloads in the trailing 30 days across the 56 laion CLIP repos (ViT-L-14-laion2B 3,950,302; ViT-B-32-laion2B 3,805,524; convnext_base_w 1,386,613). Confidence is medium rather than high only because the figure sits a few per cent above the 10M band edge and moves week to week; the warehouse sweep put it at 10,485,793, on the same side of the line. The timm and apple re-hosts of OpenCLIP checkpoints are excluded as mirrors, so this is if anything an undercount.
- https://huggingface.co/api/models?author=laion&search=CLIP&limit=100&sort=downloads&direction=-1 recorded 2026-09-11
56 laion CLIP repos with `downloads` (trailing 30 days) summing to 10,909,706.
Capability
3 high confidenceRung 3, solid workhorse: real published numbers on a named instrument, ahead of the model it reproduced, but a superseded generation - the LAION-2B checkpoints date from 2022-2023 and SigLIP 2 clears them by five points on the same measure.
- https://raw.githubusercontent.com/mlfoundations/open_clip/main/README.md recorded 2026-09-11
model table, ImageNet zero-shot acc.: ViT-H-14 LAION-2B 32B samples 78.0%; ViT-bigG-14 LAION-2B 34B samples 80.1%; ViT-L-14-quickgelu (Original CLIP) WIT 13B 75.5%.
Verified 2026-09-11