Kakao Open-Sources Kanana-2-v-decide, a Decision Model for AI Agents

Built for agent workflows

AI agents handle multi-step tasks and must keep deciding which tool to use, which information to consult and what to do next. A decision model supplies the conclusion for each step quickly so the agent can move on.

Kanana-2-v-decide is based on Kakao's multimodal model Kanana-2-v-3b. A probability shows how likely a condition is to hold. A choice gives the probability of each option and the best fit. A score gives a finer rating that reflects probabilities at each evaluation stage. In one example, a customer uploads a photo of a broken product and asks for an exchange. The model weighs the photo and the message to judge whether the item is really damaged, what type of inquiry it is and how dissatisfied the customer is. The service then routes the case to an agent or starts the exchange process. Other uses include sorting inquiries, picking tools for agents and judging product images.

The design is embedding-based: questions and options are converted into vectors and compared for similarity. This works even with many categories, and frequently used options can be stored as vectors and reused, which cuts computation and response time for repeated requests.

Speed and benchmarks

According to Kakao, the model responds in 0.04 seconds on average on a single H200 GPU, about 43 times faster than OpenAI's GPT-6 Luna. Accuracy was measured on Kakao's own Korean benchmark, Kanana Decision Bench, with 15,909 questions, and on the public Decision Index. Kakao says it scored highest on both against Microsoft's Decision-1, TypeSafe AI's Jev, Perplexity's pplx-decider-v1.1-27B and Upstage's Solar Decide.

Kakao plans to apply the model first to internal projects such as web search and automatic product categorization, refining it in live services. Longer term, it aims to merge search and decision-making into a single model.