"그 사진 찾아줘" 한마디면 끝: Kanana 멀티모달 임베딩 모델 개발기 | Model & Agent
kakao tech · 21:49 · Watch on YouTube
Kakao's team built Kana Embedding, a 2048-dim unified multimodal embedding from a vision-language model, and boosted its retrieval with gradient caching, weight merging, Matryoshka truncation, and an online-learned ha...