Kakao Open-Sources 4 Commercial 'Kanana 1.5' Language Models

Kakao has made a strong push to invigorate the domestic AI ecosystem. It has fully open-sourced four types of its self-developed artificial intelligence language model 'Kanana 1.5' under the Apache 2.0 license, which has garnered significant industry attention due to its unrestricted commercial use. The 'base' and 'instruct' models of Kanana 1.5-8b/2.1b, released via Hugging Face, have achieved a dramatic performance improvement compared to the initial models in February. Specifically, in essential capabilities for AI agent implementation such as coding, mathematical problem-solving, and function calling, they are on average 1.5 times superior to previous versions, showcasing abilities that approach the State-of-the-Art (SOTA) in the open-source domain. Another notable aspect is that while possessing competitiveness against global models, it has maintained excellent Korean language processing capabilities. Its long context understanding and ability to provide concise and accurate answers are expected to significantly enhance user experience in real service environments. Undoubtedly, the core of this open-source release is the application of the 'Apache 2.0 license'. This means that the door has been wide open for not only AI researchers and developers but also businesses to freely modify and commercially utilize Kanana 1.5. Kim Byung-hak, Kanana Performance Leader at Kakao, emphasized, "We will create a platform for growth and competition through AI technology and pursue the intrinsic value of open source." Kakao is already focusing on the development of 'Kanana 2', aiming to improve longer text processing and sophisticated inference capabilities. It remains to be seen what changes Kakao's actions will bring, as it has consistently released the Kanana series since last year, leading the development of domestic AI technology.