Lee Gyu-lee is a business writer at The Korea Times, focusing primarily on IT & telecommunications, the Ministry of Trade, Industry and Energy and KOTRA. Prior to this, she has covered a wide range of cultural news, from film, television and K-pop to lifestyle and fashion.
Kakao releases top-performing lightweight MLLM, MoE AI models
This chart shows the capability comparison between Kakao's Kanana-1.5-v-3b and Korean and international models. Courtesy of Kakao
By Lee Gyu-lee
Kakao has become the first company in Korea to make publicly available its top-performing lightweight multimodal large language model (MLLM) and mixture of experts (MoE) model.
The company announced Thursday that it released two new models — the lightweight MLLM Kanana-1.5-v-3b and the MoE-based language model Kanana-1.5-15.7b-a3b — on the global open-source artificial intelligence (AI) platform Hugging Face.
The latest release comes just two months after Kakao made available its initial set of four Kanana-1.5 models in May.
“Open-sourcing these models marks a milestone in achieving technical breakthroughs, delivering both cost efficiency and high performance,” said Kim Byung-hak, Kakao’s head of Kanana Alpha.
“It’s not just an architectural upgrade; it’s a crucial step toward product-level deployment and technological independence.”
Kanana-1.5-v-3b, which was developed from scratch entirely with Kakao’s technology, builds on the Kanana 1.5 architecture and is capable of understanding user instructions to follow and comprehend both Korean and English images. For benchmark testing in instruction-following capability, it outperformed other Korean MLLMs of similar weight by 128 percent.
Kanana-1.5-v-3b can be applied flexibly to support tasks, such as recognizing images and text, writing creative content like stories and poems, identifying important cultural landmarks or tourist sites in Korea, analyzing charts and data tables, and solving math problems.
For example, if a user uploads a photo taken near Cheonggye Stream in central Seoul, a popular tourist spot, and asks where it was taken, the model is capable of accurately identifying the site.
The company is now focused on advancing its performance with multimodal understanding, user intent following and reasoning, which are essential for agent-style AI.
The MoE model, Kanana-1.5-15.7b-a3b, was developed using an upcycling approach based on its 3 billion parameter predecessor, Kanana-Nano-1.5-3B, and is designed to deliver high-performance AI capabilities at a significantly lower cost.
Unlike typical dense models that activate all parameters for each operation, MoE only activates a subset of specialist models suited to specific tasks, leading to highly efficient computer usage at lower costs. Kakao’s MoE model only activates around 3 billion parameters out of a total of 15.7 billion during inference.
“By open-sourcing these models, Kakao is setting a new standard for the AI model ecosystem, making it easier for more researchers and engineers to access and build upon efficient, high-performing AI foundations,” Kakao said.
“The company will continue to refine these proprietary models and aims to scale up to massive, world-class flagship models — strengthening both Korea’s AI sovereignty and technical competitiveness.”
Since unveiling the original Kanana AI lineup last year, the company has continued to share new models and the development process with the public. Aside from releasing publicly four updated Kanana models in May, it also adopted the commercial-friendly Apache 2.0 license, enabling researchers and startups to freely experiment with and deploy homegrown LLMs.