Kakao Unveils Hybrid AI Model Capable of Handling Everything from Casual Conversations to Complex Reasoning

Kanana-v-4b-hybrid Performance Unveiled
Minimizing Hallucinations Through Self-Verification, Improving Accuracy
Future Model Will Automatically Switch Between General and Reasoning Modes

On January 5, Kakao announced that it has unveiled the performance of its newly developed artificial intelligence (AI) model, 'Kanana-v-4b-hybrid', which can handle everything from casual daily conversations to solving complex problems requiring logical reasoning, all within a single model.

This model was developed based on 'Kanana-1.5-v-3b', which was released as open source through the open-source AI model platform 'Hugging Face' in July of last year. Its distinguishing feature is that it goes beyond simply converting images to text or describing them; it synthesizes information, performs calculations, and conducts self-verification through a self-checking process.

Kakao announced on the 5th that it has unveiled the performance of its newly developed artificial intelligence (AI) model, 'Kanana-v-4b-hybrid'. Photo by Kakao

Through this process, the AI model minimizes hallucinations and reduces calculation errors and omissions of conditions that can easily occur in complex problems such as tables, receipts, or math questions, thereby improving accuracy.

To achieve the high performance of the Kanana-v-4b-hybrid model, Kakao implemented a four-stage training procedure: foundational training, long chain-of-thought (Long CoT), offline reinforcement learning, and online reinforcement learning.

The model has also demonstrated competitiveness in logical reasoning in Korean. It was trained to directly understand and reason with Korean questions, and, compared to existing global models that translate Korean questions into English for reasoning and then retranslate the answers, it excels in both context and logic.

Based on this Korean language performance, the model achieved high accuracy rates on questions from the College Scholastic Ability Test (CSAT) in subjects such as social studies and mathematics, accurately understanding the Korean context. In the KoNET (Korea National Educational Test Benchmark), an AI academic evaluation benchmark based on the Korean education system, it scored 92.8 points.

In performance evaluations against similarly sized global models such as Qwen3-VL-4B, InternVL3.5-4B, and GPT-5-nano, the Kanana-v-4b-hybrid surpassed them in areas requiring complex reasoning, such as mathematics and science, as well as in general visual comprehension capabilities.

Going forward, Kakao plans to further advance the model so that users will not need to select a specific model themselves; the AI will automatically determine the complexity of the question and switch between general and reasoning modes as needed. This will enable a cost-effective model that provides a seamless user experience for both simple questions and complex analytical requests within a single chat window.

Kim Byunghak, Performance Leader of Kakao Kanana, stated, "Kanana-v-4b-hybrid is a model that can think and respond most naturally and accurately in the Korean language environment, making it an innovative research achievement that allows both everyday and complex tasks to be entrusted to a single AI. Through the development of a proprietary AI model with high performance and efficiency specialized for Korean, we aim to enhance our global competitiveness and continue to lead the advancement of the domestic AI ecosystem."