Kakao Unveils Next-Generation Language Model “Kanana-2” on the 19th

Kakao has launched its next-generation language model, “Kanana-2,” on Hugging Face as an open-source tool. Following the success of the “Kana” series, this advanced model boasts enhanced performance and efficiency, particularly in AI applications that require understanding user commands. The release features three types: Base, Instruct, and Thinking-specific models, with the inference model being a notable highlight, allowing developers to fine-tune using their own data. Kanana-2 significantly improves tool-calling and instruction-following capabilities, exceeding performance metrics of its predecessor, Kana-1.5. This model supports six languages, including Japanese and Vietnamese, and implements cutting-edge techniques like “MLA” and “MoE” for efficient long-input processing. Kakao aims to expand the model’s size and capabilities, focusing on complex AI scenarios. With an emphasis on performance and openness, Kakao seeks to enhance the AI research ecosystem globally through practical AI solutions.

Source link