Kakao has released four open-source lightweight language models (SLM) developed under its artificial intelligence (AI) brand, Kanana. This initiative aims to enhance AI technology competitiveness and contribute to the domestic ecosystem.
On July 28, Kakao announced the open-source release of four lightweight language models that can operate on devices such as smartphones. The models are part of the Kanana-2 series and include: △Kanana-2-1.3B-base △Kanana-2-1.3B-instruct △Kanana-2-3B-base △Kanana-2-3B-instruct.
Kakao explained that it has been continuously researching lightweight language models that can deliver service-level performance despite their smaller size. This release reflects the growing importance of lightweight models that can run directly on devices like smartphones and personal computers, amid a competitive landscape focused on large language models (LLMs). Global companies are also prioritizing on-device AI as a key strategy, with many releasing lightweight open-source models to establish a foothold in the lightweight AI ecosystem.
The models released are based on the technology and experience accumulated during various model development processes. Currently, Kakao's lightweight models are utilized in services such as 'Kanana in KakaoTalk,' 'KakaoTalk Conversation Summary and Call Summary,' and 'AI National Secretary.'
Despite their small size, the models demonstrate global-level performance in both Korean and English. They have outperformed most representative benchmarks, including Korean, English, knowledge, mathematics, and code, compared to similar-sized state-of-the-art open-source models. Notably, the Kanana-2-1.3B-instruct model, released as open-source, along with the smaller Kanana-2-0.9B-instruct, shows competitive performance in real service applications, including conversation, knowledge, code, mathematics, instruction execution, and tool invocation, rivaling global models like Qwen and Gemma.
Additionally, Kakao has applied its self-developed Korean-specific tokenizer, improving Korean text processing efficiency by over 30%. The tokenizer is a key technology that determines how an AI model breaks down text into smaller units. By using a tokenizer optimized for Korean, the same sentence can be processed in fewer units, resulting in faster computation speeds and reduced costs. Kakao has implemented this Korean-specific tokenizer across all models to enhance performance and cost competitiveness.
The models are designed to be optimized for on-device environments, such as smartphones, where memory and computational resources are limited. To address the issue of rapidly increasing memory usage during long conversations, a 'sliding window attention' structure has been introduced. This allows for stable performance while reducing memory usage by up to 72.7% for conversations of up to 32K tokens (approximately 24,000 words).
Kakao will distribute these models under the 'Kanana Open License,' allowing commercial use. Developers, startups, and research institutions can utilize the technology without restrictions, contributing to the expansion of high-performance Korean AI model applications and the activation of new service development in the country.
Noh Byeong-seok, leader of Kakao's Unified Foundation Model Performance, stated, "As we prepare for the era of agentic AI, we recognize the importance of both large-scale AI in the cloud and lightweight AI on devices. Kakao is enhancing its technological competitiveness in both areas, and we hope that these open-source models will lead to the development of new AI services by more developers and companies, contributing to the activation of the domestic AI ecosystem."
Kanana is Kakao's integrated AI brand, encompassing not only language models but also multimodal, voice, and image generation models, along with related services and development organizations. Kakao has progressively advanced its models, starting with the lightweight Kanana Nano in 2025, followed by Kanana 1.5, which enhances coding, mathematics, and tool invocation capabilities, and Kanana-2, which focuses on agentic AI.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.