Kakao Kanana-2: Tops Google, Alibaba in AI Safety

Kakao's Kanana-2 series of small language models scored higher on AI safety than Google's Gemma and Alibaba's Qwen. This was based on tests using Kakao's own AI safety evaluation platform. The platform measured toxicity and bias in two models, Kanana-2-1.3B-Instruct and Kanana-2-3B-Instruct. Both models posted overall scores ahead of the global open-source competitors. For example, Kanana-2-1.3B-Instruct scored 0.70, compared to Gemma's 0.68 and Qwen's 0.58. The tests used AssurAI, a Korean-language safety benchmark built by the Telecommunications Technology Association, KAIST, and Kakao. AssurAI covers 9,560 evaluation items across 35 risk categories, regrouped by Kakao into five broad areas. Kakao plans to expand this safety testing to all future self-developed AI models. This matters because it shows a commitment to rigorous safety checks for AI technology before it reaches the public.