Kakao Unveils Hybrid AI Model Combining General Conversation and Inference

Kakao has introduced a new AI model capable of both general conversation and complex reasoning. On the 5th, Kakao revealed the performance of its self-developed multimodal language model, ‘Kanana-v-4b-hybrid’. This model is based on the ‘Kanana-1.5-v-3b’ that was released as open-source in July of last year. The most notable feature is that the AI undergoes a ‘self-check’ process, wherein it verifies its own calculations during information synthesis and computation. The company explained that through this, it has reduced hallucinations and improved accuracy in complex tasks such as analyzing tables and receipts and solving math problems. The training process involves a four-step procedure: basic training, Long Chain of Thought (CoT), offline reinforcement learning, and online reinforcement learning. Notably, it was trained to understand and think in Korean without translating questions into English, thus maintaining its contextual understanding capability. Regarding performance metrics, Kakao stated that it scored 92.8 points on ‘KoNET’, an AI academic assessment benchmark based on the Korean education system. Additionally, it demonstrated high performance in the areas of science, engineering, and document understanding in comparative evaluations with similarly sized domestic and international models like Qwen3-VL-4B, InternVL3.5-4B, and GPT-5-nano. Kakao plans to add a feature that allows the AI to automatically switch between ‘general mode’ and ‘reasoning mode’ based on the complexity of the questions. This is intended to enhance resource utilization efficiency. Kim Byung-hak, the Kanana performance leader, stated, “This is a model that can respond naturally in a Korean language environment,” and “We have advanced our research results to allow one AI to handle both everyday conversations and complex tasks.”