Kakao Releases 'Kanana-v-embedding', a 2B Parameter Multimodal Embedding Model, Outperforming Jina v4 in Korean Performance

The Kakao Kanana organization has developed 'Kanana-v-embedding', a multimodal embedding model that simultaneously understands text and images. This model features a 2-billion (2B) parameter scale and recorded significant performance advantages over the existing large model Jina Embeddings v4 (3.8B) in key tasks such as Korean sentence search (KoEmbed-Sent). It applies Matryoshka Representation Learning techniques to support flexible embedding dimensions ranging from 64 to 2,048, and utilizes state-of-the-art training techniques such as gradient caching and hard negative mining. Additionally, the organization constructed the dataset 'KoEmbed', optimized for Korean service environments, for training purposes. The model is currently applied to the company's internal advertising review platform and is in active use for similar material search functionality.