SEOUL, July 24 (AJP) - Kakao has released two new open-source artificial intelligence models, including Korea’s first publicly available mixture-of-experts (MoE) model and a lightweight multimodal model.
The company announced on Thursday that both models — Kanana-1.5-v-3b, a compact multimodal language model, and Kanana-1.5-15.7b-a3b, an MoE model designed for efficient inference — are now available on Hugging Face, a leading platform for open-source AI tools.
The release underscores Kakao’s ambitions to move beyond consumer messaging and establish itself as a serious player in advanced AI research and development.
“As a developer of in-house models and a provider of large-scale digital services like KakaoTalk, we see it as our role to help strengthen Korea’s AI infrastructure,” the company said in a statement.
Multimodal AI models are capable of processing various forms of input — text, images, and sometimes audio — simultaneously.
The Kanana-1.5-v-3b model, optimized for deployment in mobile and IoT environments, can understand images and follow complex user instructions in both Korean and English.
Built entirely with Kakao’s own technology, the model delivers performance that the company says rivals larger international offerings — achieving up to 128 percent higher instruction-following accuracy than comparable open-source Korean multimodal models.
Copyright ⓒ Aju Press All rights reserved.
View more comments