NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI
PyPI, #2942 most downloaded on PyPI
Open reproduction of contrastive language-image pretraining (CLIP) and related.
Latest stable
4.2.2
Last release 12 days ago
10 Nov 2026
Release timing varies
gaps range from 2 weeks to 8 months
Most releases are documented
notes for 42 of 57 stable releases
Nothing withdrawn
no release was ever pulled
5 years old
59 releases · first in 2022
Drag across the ruler to compare, or tap a release to open it.
The first 4.x release on PyPI, and it adds official pretrained NaFlex CLAP audio-text models. It also moves training to a new stack and adds several m
The first 4.x release on PyPI, and it adds official pretrained NaFlex CLAP audio-text models. It also moves training to a new stack and adds several model families. Earlier 4.x version numbers were development versions only.
Pretrained image-text inference is intended to remain compatible, but training scripts and downstream integrations need review. See the migration guide. To stay on 3.x, run pip install 'open_clip_torch<4'.
coca2_* configsTrainingTask wrappers, dict-based batches, FSDP2 (--fsdp), torch.compile strategies, and length bucketingweights_only=True, and OpenAI weights now load from the Hugging Face Hub
instead of JIT archivestorch>=2.6 (was >=2.0)--precision is now amp_bf16 (was amp), and SigLIP's --loss-dist-impl now defaults to gather--horovod, --torchscript, and --trace are removedload_openai_model, list_openai_models, build_model_from_openai_state_dict, and trace_model are removed; use
create_model_from_pretrained(..., pretrained='openai') for OpenAI weightsTrainingTask instead of a separate model and losspython -m open_clip_train.legacy_main remains available as a compatibility path for older image-text training scriptsSee the migration guide for the full list of changes.
Full Changelog: https://github.com/mlfoundations/open_clip/compare/v3.3.0...v4.2.2
Your coding agent can read these notes before it upgrades. Set up the MCP server →
The first 4.x release on PyPI, and it adds official pretrained NaFlex CLAP audio-text models. It also moves training to a new stack and adds several m
The first 4.x release on PyPI, and it adds official pretrained NaFlex CLAP audio-text models. It also moves training to a new stack and adds several model families. Earlier 4.x version numbers were development versions only.
Pretrained image-text inference is intended to remain compatible, but training scripts and downstream integrations need review. See the migration guide. To stay on 3.x, run pip install 'open_clip_torch<4'.
coca2_* configsTrainingTask wrappers, dict-based batches, FSDP2 (--fsdp), torch.compile strategies, and length bucketingweights_only=True, and OpenAI weights now load from the Hugging Face Hub
instead of JIT archivestorch>=2.6 (was >=2.0)--precision is now amp_bf16 (was amp), and SigLIP's --loss-dist-impl now defaults to gather--horovod, --torchscript, and --trace are removedload_openai_model, list_openai_models, build_model_from_openai_state_dict, and trace_model are removed; use
create_model_from_pretrained(..., pretrained='openai') for OpenAI weightsTrainingTask instead of a separate model and losspython -m open_clip_train.legacy_main remains available as a compatibility path for older image-text training scriptsSee the migration guide for the full list of changes.
Full Changelog: https://github.com/mlfoundations/open_clip/compare/v3.3.0...v4.2.1
First 4.x release on PyPI. See the migration guide for breaking changes when upgrading from 3.x.
TrainingTask stack with dict-based batches, FSDP2, and torch.compile strategiesweights_only=TrueReleases between 2.24.0 and 4.2.1 are not recorded in this file; see the GitHub releases.
Nothing published for this version
Nothing published for this version
Nothing published for this version
Remove non-existent MetaCLIP 2 L/14 checkpoint by @voidism in https://github.com/mlfoundations/open_clip/pull/1111
Full Changelog: https://github.com/mlfoundations/open_clip/compare/v3.1.0...v3.2.0
Add support for MetaCLIP2 WorldWide models by @rwightman in https://github.com/mlfoundations/open_clip/pull/1100
Full Changelog: https://github.com/mlfoundations/open_clip/compare/v3.0.0...v3.1.0
Initial work on adding local-dir: schema for model & tokenizer loading from local folder by @rwightman in https://github.com/mlfoundations/open_clip/p
Full Changelog: https://github.com/mlfoundations/open_clip/compare/v2.32.0...v3.0.0
API for getting intermediate image and text features, forward_intermediates() by @rwightman in https://github.com/mlfoundations/open_clip/pull/1035
Full Changelog: https://github.com/mlfoundations/open_clip/compare/v2.31.0...v2.32.0
Add SigLIP2 models by @rwightman in https://github.com/mlfoundations/open_clip/pull/1033
Full Changelog: https://github.com/mlfoundations/open_clip/compare/v2.30.0...v2.31.0
Support using timm optimizers for alternative to adamw default by @rwightman in https://github.com/mlfoundations/open_clip/pull/979
Full Changelog: https://github.com/mlfoundations/open_clip/compare/v2.29.0...v2.30.0
All default pretrained weights pushed to HF hub by @rwightman in https://github.com/mlfoundations/open_clip/pull/970
Full Changelog: https://github.com/mlfoundations/open_clip/compare/v2.28.0...v2.29.0
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Fix missing space in error message
Stop checking model name when loading CoCa models
Nothing published for this version
* Add EVA models * Support serial worker training * Fix Python 3.7 compatibility
* Add DataComp models
Enable int8 inference without .weight attribute
.weight attribute* Update push_to_hf_hub
Nothing published for this version
Fixes for context_length and vocab_size attributes
Fixes for context_length and vocab_size attributes
* Add improved g-14 weights * Update protobuf version
Add samples per second per gpu logging
Move dataset mixtures logic to shard level
Add support for dataset mixtures with different sampling weights
Updated convnext configs for consistency
Add MSCOCO CoCa finetunes to pretrained models
* coca support and weights * ConvNeXt-Large weights
hf-hub:org/model_id support for loading models w/ config and weights in Hugging Face Hub
hf-hub:org/model_id support for loading models w/ config and weights in Hugging Face HubAdded an up-to-date example slurm script for large training jobs.
base & base_w pretrained models addedtimm- model prefix removed from configstimm augmentation + regularization (dropout / drop-path) supportedFix wandb collapsing multiple parallel runs into a single one
Fix braceexpand memory explosion for complex webdataset urls
* Fix release
wrapped patchdropout in a torch.nn.Module
override the default patch dropout value in 'vision_cfg'
add support for gradient accumulation
add multilingual H/14 xlm roberta large
* fix setup.py _read_reqs
pretrained B/32 xlm roberta base: first multilingual clip trained on laion5B
Add missing hf_tokenizer_name in CLIPTextCfg.
Fix #211, missing RN50x64 config. Fix type of dropout param for ResNet models
Implement grad checkpointing for hf model.
Generalizable Text Transformer with HuggingFace Models (@iejMac)
Add checksum verification for pretrained model weights
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
add missing openai RN50x64 model
* ViT-B/16+ * Add grad checkpointing support * more robust data loader
Nothing published for this version
Nothing published for this version
Nothing published for this version
One column per quarter.