KaKao

- Open-source release of four Kanana-2 SLMs optimized to run on devices such as smartphones

- Performance on par with global models of comparable size, with efficiency improved more than 30% through a proprietary tokenizer

 

- Free access to lightweight AI models for developers, research institutions, and businesses under a license that permits commercial use

 

[July 28, 2026] Kakao (CEO Shina Chung)  announced that it is open-sourcing small language models (SLMs) developed in-house as part of its Kanana AI model family, bolstering its AI technology capabilities and contributing to the domestic AI ecosystem.

 

Kakao released four SLMs on Hugging Face on July 28, making them available as open-source models optimized to run on devices such as smartphones. The four models in the Kanana-2 SLM series are Kanana-2-1.3B-base, Kanana-2-1.3B-instruct, Kanana-2-3B-base, and Kanana-2-3B-instruct.

 

As competition over the performance of large language models (LLMs) intensifies, importance of SLMs that can run directly on personal devices such as smartphones and PCs is increasingly being emphasized in the AI industry. Global companies are also seeking to establish a foothold in the lightweight AI ecosystem by making on-device AI a key strategic focus and successively releasing lightweight open-source models.

 

Amid this trend, Kakao has continued to research SLMs that deliver performance sufficient to support real-world services despite their compact size. The newly released models were developed based on technologies and expertise accumulated through various model development projects. Kakao’s SLMs are currently being used across a range of services, including Kanana in KakaoTalk, KakaoTalk’s chat and call summary features, and AI MOIS e-Gov.

 

Despite their compact size, the newly released models deliver global-level performance in both Korean and English. Compared with the latest SOTA open-source models of similar size, they demonstrated strong performance across most major benchmarks, including Korean and English language comprehension, knowledge, mathematics, and coding. In particular, the open-sourced “Kanana-2-1.3B-instruct” and the smaller “Kanana-2-0.9B-instruct,” which was developed alongside the released models, demonstrated competitive performance against comparable global models such as Qwen and Gemma in practical service use, including conversation, knowledge, coding, mathematics, instruction following, and tool calling.

 

The models also incorporate Kakao’s proprietary Korean-specialized tokenizer, improving Korean processing efficiency by more than 30% compared with previous approaches. A tokenizer is a core technology that determines how text is divided into smaller units for processing by an AI model. The more optimized a tokenizer is for Korean, the fewer units are required to process the same text, resulting in faster computation and lower costs. Kakao applied its Korean-specialized tokenizer, developed in-house last year, across the newly released models, enhancing both performance and cost efficiency.

 

The models are also designed and optimized for on-device environments with limited memory and computing resources, such as smartphones. To address the sharp increase in memory usage when processing long conversations, Kakao adopted a “Sliding Window Attention” structure. This allows the models to maintain stable performance while reducing memory usage by up to 72.7% when processing conversations of up to 32K tokens (approximately 24,000 words).

 

Kakao is releasing the models under the “Kanana Open License,” which permits commercial use. Developers, startups, and research institutions can use the models without additional restrictions, helping broaden access to high-performance Korean AI models in Korea and accelerate the development of new AI services.

 

Byungseok Roh, Head and Executive Lead of Unified Foundation Model Studio at Kakao, 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 strengthening its technological capabilities in both areas, and we hope that the open-source models released today will lead to the development of new AI services by more developers and businesses—ultimately contributing to the growth of Korea’s AI ecosystem.” (E.O.D.)

 

Reference 1: Hugging Face https://huggingface.co/collections/kakaocorp/kanana-2-slm