Kakao Brain releases open-source ViT and ALIGN models trained on COYO-700M dataset

AI Impact Summary

Kakao Brain released open-source ViT and ALIGN models trained on the COYO-700M dataset, marking the first public ALIGN release and the first ViT/ALIGN models trained on an open dataset. The team asserts parity with Google's ViT/ALIGN on core benchmarks and even improvements on MS-COCO retrieval, with demos hosted on Hugging Face for immediate experimentation. COYO provides richer metadata (aesthetic scores, watermark, face counts) and supports streaming access via the 🤗 Datasets library, enabling reproducible research and scalable evaluation for image-text tasks like retrieval and captioning. This release lowers barriers for teams to prototype, reproduce results, and compare cross-modal architectures without requiring access to private datasets, though licensing and compute considerations remain important for production use.