KaKao

- Four Kanana 1.5 models released on Hugging Face, improving on February’s version; commercial use permitted under Apache 2.0 license

- Focused on strengthening AI agent capabilities, delivering global-level performance in coding, mathematics, and function calling

- Commercial license marks a key milestone for securing technology leadership and fostering a collaborative AI ecosystem

 

[May 23, 2025] Kakao Corporation continues to advance AI ecosystem expansion and technology accessibility by open-sourcing its in-house developed AI models.

 

Kakao (CEO Shina Chung) announced on May 23 the release of four models from its Kanana 1.5 language model lineup—including 8B and 2.1B parameter versions—on Hugging Face under the Apache 2.0 license, enabling free modification and commercial use.

 

This release features significant performance upgrades over the February version, with four models in its lineup: Kanana-1.5-8b-base; Kanana-1.5-8b-instruct; Kanana-1.5-2.1b-bas; and Kanana-1.5-2.1b-instruct.

 

Kanana 1.5 emphasizes enhanced functionalities essential for building agentic AI. It maintains superior Korean language performance compared to global counterparts, achieving a 1.5-fold improvement in coding, mathematical problem-solving, and function calling capabilities, reaching state-of-the-art (SOTA) open-source model performance.

 

Further improvements include more effective handling of long-context inputs and providing concise, optimized responses. The models support more stable processing of longer texts and improve user experience by balancing response accuracy with brevity in service environments.

 

Kakao is also developing the next-generation Kanana 2 model, aiming to evolve in areas such as longer input handling, refined reasoning, and structural efficiency.

 

With this open-source release, Kakao strives to make consistent contributions to the revitalization of Korea’s large language model (LLM) ecosystem. By applying a commercial license, Kakao empowers AI researchers, developers, and companies to freely tune and deploy the models tailored to their specific needs, strengthening technological competitiveness while fostering a collaborative AI ecosystem.

 

Kakao Kanana Performance Lead Byung-hak Kim commented, “Through this open-source release, we strive to build an environment where growth and competition in AI technology coexist, creating ongoing opportunities for technological advancement within the ecosystem. We will pursue both the continuous improvement of AI model performance and the fundamental values of open source.”

 

Kakao revealed its proprietary AI model lineup “Kanana” last year and shared the performance and development of language models, multimodal language models, and visual generation models through its official tech blog. In February, Kakao released the Kanana Nano 2.1B model on Hugging Face and published a technical report on the Kanana language model research on arXiv. More recently, Kakao unveiled Kanana-o, Korea’s first integrated multimodal language model capable of simultaneously processing text, speech, and images. (E.O.D.)

 

[Reference]

* Open-source release of four Kanana models on Hugging Face: https://kko.kakao.com/kananallm

* Official Kakao Tech blog

- Kakao’s smarter language model Kanana 1.5: https://tech.kakao.com/posts/706

- Development of Kanana 1.5v: https://tech.kakao.com/posts/707