mlfoundations/open_clip

Supports diverse Vision Transformer and ConvNet architectures trained on large-scale datasets (LAION-2B, DataComp-1B) with published scaling laws, achieving competitive zero-shot ImageNet accuracy up to 85.4%. Integrates with PyTorch, Hugging Face model hub, and timm for image encoders, enabling efficient embedding computation via the clip-retrieval library. Offers flexible model loading from local checkpoints or HuggingFace, with pre-trained weights optimized for both inference and fine-tuning workflows.

13,496 stars and 2,903,706 monthly downloads. Used by 18 other packages. Actively maintained with 1 commit in the last 30 days. Available on PyPI.