Kakao Releases Four Open-Source Lightweight Kanana Models Specialized for Korean

Four Kanana-2 SLM Series Models Released
Optimized for smartphones and other devices... Already in use in services such as "Kanana in KakaoTalk"
Performance comparable to similar-sized global models

Kakao announced on July 28, 2026, that it has open-sourced the lightweight language model (SLM) of its independently developed artificial intelligence (AI) model, "Kanana."

The Kanana-2 lightweight language model series that Kakao released on the open-source AI platform Hugging Face includes Kanana-2-1.3B-Base, Kanana-2-1.3B-Instruct, Kanana-2-3B-Base, and Kanana-2-3B-Instruct. These models have been downsized to enable operation on devices such as smartphones.

Kakao announced on the 28th that it has open-sourced the lightweight language model (SLM) of its self-developed artificial intelligence (AI) model "Kanana." Performance indicators of Kanana lightweight language model. Kakao

In response to the recent trend highlighting the importance of lightweight language models that can run on personal devices, Kakao has been researching compact models with performance sufficient for practical service applications. The released models are already being used in a wide range of services, including "Kanana in KakaoTalk," "KakaoTalk Conversation and Call Summaries," and "AI National Secretary."

According to Kakao, the models released this time achieve global-level performance in both Korean and English. In particular, Kanana-2-1.3B-Instruct and Kanana-2-0.9B-Instruct have demonstrated competitive performance—rivaling global models such as Qwen (Alibaba) and Gemma (Google)—in practical service capabilities like conversation, knowledge, coding, mathematics, instruction following, and tool use.

Kakao has also improved Korean language processing performance. By applying its self-developed Korean-specialized tokenizer, Kakao has enhanced the efficiency of Korean language processing by more than 30% compared to previous systems. Since the tokenizer is the technology that determines the unit for breaking down a sentence when an AI model processes text, using a tokenizer optimized for Korean allows sentences to be handled with fewer units, resulting in faster computation speeds and reduced costs.

The models are specifically designed to be optimized for on-device environments, such as smartphones, where computing power is limited. To address the increased memory usage when processing long conversations, Kakao introduced a "sliding window attention" structure, allowing stable performance while reducing memory usage by up to 72.7% during conversations as long as 32K (32,000 tokens).

Kakao is distributing these models under the "Kanana Open License," which permits commercial use.

Byeongseok Noh, Unified Foundation Model Performance Leader at Kakao, said, "As we prepare for the era of agentic AI, we recognize the importance of both large-scale cloud-based AI and lightweight on-device AI. Kakao is strengthening its technological competitiveness in both areas, and we hope that releasing these open-source models will encourage more developers and companies to develop new AI services, vitalizing Korea's AI ecosystem."

This content was produced with the assistance of AI translation services.