Kakao Kanana-2 Beats Global AI in Safety Assessments

Kakao's self-developed language model, Kanana-2, has outperformed major global models in safety assessments. Here's the thing: Kakao released results for two models in its open-source Kanana-2 series, showing they scored higher than comparable models from Google and Alibaba. What's interesting is that these lightweight models, Kanana-2-1.3B-Instruct and Kanana-2-3B-Instruct, both achieved an overall score of 0.70. This beat Google's Gemma, which scored 0.68 and 0.66, and Alibaba's Qwen, with scores of 0.58 and 0.62. The assessments used AssurAI, a benchmark designed for South Korea's social and cultural context, measuring 9,560 evaluation items across 35 risk categories. The 1.3B model showed strength in crime and illegal activities, while the 3B model excelled in sexual content, child protection, and rights infringement. Kakao plans to expand these safety assessments to all its AI models, moving beyond text to multimodal and agentic AI. The bottom line: this demonstrates a commitment to responsible AI development and could lead to safer AI applications.