Kakao Releases Four Versions of Its Proprietary AI Model 'Kanana 1.5' as Open Source

Anyone Can Modify and Use Commercially
Development of 'Kanana 2' Underway

Kakao's proprietary language model 'Kanana' is newly introduced as the upgraded 'Kanana 1.5' and released as open source just three months after its initial unveiling. Photo by Kakao

Kakao has released four versions of its proprietary language model 'Kanana (Kanana)' 1.5 as open source to promote the expansion of the artificial intelligence (AI) ecosystem.

On May 23, Kakao announced that it had released the 8B (8 billion parameters) and 2.1B (2.1 billion parameters) models from the Kanana 1.5 lineup via Hugging Face. The four models released are Kanana-1.5-8b-base, Kanana-1.5-8b-instruct, Kanana-1.5-2.1b-base, and Kanana-1.5-2.1b-instruct. They are licensed under Apache 2.0, allowing anyone to freely modify and use them for commercial purposes.

Kanana 1.5, which has been upgraded from the previous model unveiled in February, focuses on enhanced features for implementing agentic AI. While maintaining outstanding Korean language performance, it has achieved an average performance improvement of 1.5 times compared to its predecessor in coding, solving mathematical problems, and function calling capabilities.

Additionally, it offers improved usability through better understanding of long contexts and more concise answers. The model can now stably process longer text inputs than before, and it enhances user-perceived performance in service environments by providing responses optimized for both accuracy and length.

Kakao is currently developing an even more advanced version, Kanana 2. The new model aims to evolve in several areas, including handling longer inputs, more sophisticated reasoning, and structural efficiency.

With this open-source release, Kakao plans to continue contributing to the activation of the domestic large language model (LLM) ecosystem. By applying a commercial license that allows AI researchers, developers, and companies to freely tune and utilize the models for their own purposes, the company aims to secure technological competitiveness and build a collaborative AI ecosystem at the same time.

Kim Byunghak, Kanana Performance Leader at Kakao, said, "Through this open-source release, we aim to create an environment where growth and competition through AI technology can coexist, and to continuously provide opportunities to advance technology within this ecosystem. We will pursue not only the advancement of AI model performance but also the fundamental value of open source."