Kakao Open-Sources Kanana-2-v-Decide Multimodal Decision Model

In one example, a buyer asks for an exchange and attaches a photo of a damaged product. The model assesses whether the item is broken, what type of inquiry it is and the customer's sentiment. A service can then follow preset rules to explain exchange options or pass the case to a human agent.

Built for speed

To cut latency, the model converts questions and answer options into numeric vectors and compares them. Vectors for options can be stored and reused, which reduces computation when the same request recurs and suits tasks with many categories. According to Kakao, average response time on a single Nvidia H200 GPU was under 0.04 seconds, 43 times faster than GPT-6 Luna.

Kakao also says the model scored highest among the compared models on Kanana Decision Bench, its own Korean benchmark of 15,909 questions, and on the public Decision Index. The comparison included Microsoft's Decision-1, Typesafe AI's Zeb, Perplexity's pplx-decider-v1.1-27B and Upstage's Solar Decide. Typesafe AI, a US startup, unveiled Zeb in September, a sign that decision-focused models are gaining ground as components for agents that must pick tools, information and next steps.

The model is released under the Kanana Open License, which allows commercial use as well as research, so developers and companies in Korea and abroad can use it to build AI agent services.

Kakao plans to deploy the technology first in internal web search and product classification, refine it through real-world use and, over the longer term, merge search and decision-making functions.