[Edaily Reporter Lee So-Hyun ] Kakao(035720)has released its proprietary lightweight language model (SLM), which can run directly on devices such as smartphones and personal computers, as open source. As the competition in artificial intelligence (AI), which has centered on large language models (LLMs), expands to on-device AI, the company is setting out to expand the technology ecosystem with a lightweight model that excels in Korean.

Kakao announced on the 28th that it has released four lightweight language models from its in-house AI model series “Kanana” on the global AI platform Hugging Face.

The models released this time are part of the Kanana-2 lightweight language model series. Specifically, they include four variants: △Kanana-2-1.3B-base, △Kanana-2-1.3B-instruct, △Kanana-2-3B-base, and △Kanana-2-3B-instruct.

Recently, in the AI industry, the importance of lightweight models—which can run directly on devices with limited computing power, such as smartphones and personal computers—has been growing alongside that of cloud-based large models. Global companies are also positioning on-device AI as a key strategy or releasing lightweight open-source models one after another to secure a foothold in the related ecosystem.

Amid this trend, Kakao has been researching lightweight language models that, despite their small size, are capable of being applied to real-world services. Currently, Kakao’s lightweight models are being utilized in various services, including “Kanana in KakaoTalk,” “KakaoTalk Conversation and Call Summarization,” and “AI National Secretary.”

The newly released models are notable for delivering world-class performance in both Korean and English despite their small size. The Kanana-2 series outperformed the latest open-source models of similar size on key benchmarks, including Korean, English, general knowledge, mathematics, and code.

In particular, the company explained that the smaller Kanana-2-0.9B-instruct—developed alongside the open-source Kanana-2-1.3B-instruct—demonstrated competitive performance against global models of the same class, such as Qwen and Gemma, in real-world service applications including conversation, knowledge, code, math, instruction execution, and tool invocation.

Kakao also improved the efficiency of Korean language processing by applying its proprietary Korean-specialized tokenizer across the entire model. A tokenizer is a core technology that determines the units into which sentences are divided when an AI model processes text. Applying a tokenizer optimized for Korean allows the same sentence to be processed into fewer units, thereby increasing processing speed and reducing costs. Kakao announced that by applying the Korean-specific tokenizer—which was completed last year—it has improved Korean processing efficiency by more than 30% compared to previous versions.

The design also takes on-device environments into account. Kakao introduced a “sliding window attention” structure to ensure stable processing of long conversations even in environments with limited memory and computational resources, such as smartphones. The company explained that this allows the model to maintain stable performance while reducing memory usage by up to 72.7% even in conversations up to 32K—approximately 32,000 tokens—in length.

This model is being released under the “Kanana Open License,” which permits commercial use. As a result, developers, startups, research institutions, and companies can utilize Kakao’s lightweight AI model at no additional cost. Kakao expects this to contribute to expanding the adoption of high-performance Korean AI models in South Korea and stimulating the development of new services.

Noh Byung-seok, Performance Leader for Kakao’s Unified Foundation Model, stated, “As we prepare for the era of agentic AI, we are realizing the importance of both large-scale AI in the cloud and lightweight on-device AI.” He added, “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, thereby contributing to the revitalization of the domestic AI ecosystem.”