Google’s ViT and ALIGN models have been something of a tease for the research community. The architectures were public, the papers were detailed, but the training data — 300 million images for ViT, 1.8 billion image-text pairs for ALIGN — stayed locked away. That made replication impossible. Kakao Brain just changed the equation.
The Korean AI lab released open-source versions of both models, trained entirely on COYO-700M, a public dataset of 700 million image-text pairs scraped from webpages. Same architectures, same hyperparameters as Google’s originals. The kicker: Kakao’s ALIGN-B7-Base, trained on less than 40% of the data Google used, performs on par with Google’s version on Image KNN classification and actually beats it on MS-COCO retrieval tasks in both directions. Their ViT-L/16 matches Google’s numbers on ImageNet and ImageNet-ReaL at 384 and 512 resolutions.
The COYO dataset itself deserves attention. It occupies a middle ground between LAION-2B and Google’s private ALIGN corpus — smaller than LAION but richer in metadata. Every pair comes with aesthetic scores, more robust watermark scores, and face count data. That’s genuinely useful for researchers who want finer control over what they train on. LAION gives you volume; COYO gives you volume plus metadata. For a lot of use cases, that’s the better trade.
This isn’t just another model dump. It’s the first open-source ALIGN implementation ever, and the first time both ViT and ALIGN are reproducible end-to-end — code, weights, and training data all public. The demos are already live on Hugging Face. After years of watching Google publish impressive results on private datasets, researchers finally have a transparent baseline that actually matches the state of the art.
