Kakao unveils 4 Kanana-2 SLMs, claims performance on par with global rivals

Kakao has unveiled four small language models (SLMs) in its proprietary Kanana series, saying the models deliver performance comparable to similarly sized global counterparts despite their compact footprint.

Kakao said Tuesday it released four models in the Kanana-2 SLM series on Hugging Face: Kanana-2-1.3B-base, Kanana-2-1.3B-instruct, Kanana-2-3B-base and Kanana-2-3B-instruct.

The newly released models were developed drawing on the technology and experience Kakao has accumulated across various model development projects. The company's SLMs are already in use across several services, including "Kanana in KakaoTalk," KakaoTalk's conversation and call summarization features, and the AI National Secretary service.

Despite their small size, the models achieve global-level performance in both Korean and English. Across most major benchmarks — covering Korean, English, knowledge, mathematics and code — they outperformed state-of-the-art open-source models of comparable size.

Both the open-source Kanana-2-1.3B-instruct and the smaller Kanana-2-0.9B-instruct demonstrated competitive performance against global models of the same class, such as Qwen and Gemma, across practical service capabilities including conversation, knowledge, code, mathematics, instruction-following and tool calling.

The models also incorporate a Korean-specialized tokenizer developed in-house by Kakao, improving Korean text processing efficiency by more than 30 percent compared with previous approaches. A tokenizer is a core technology that determines how an AI model breaks sentences into units when processing text. The more optimized a tokenizer is for Korean, the fewer units it needs to handle the same sentence — speeding up computation and reducing costs.

The models are also designed for on-device environments with limited memory and computing resources, such as smartphones. This allows them to cut memory usage by up to 72.7 percent while maintaining stable performance in conversations of up to 32K tokens — roughly 24,000 words.

In addition, Kakao is distributing the models under the "Kanana Open License," which permits commercial use. Developers, startups, research institutions and others can use the technology without restriction.

"We hope the release of these open-source models will lead more developers and companies to build new AI services, contributing to the vitalization of the domestic AI ecosystem," said Noh Byeong-seok, Kakao's unified foundation model performance leader.