Kakao releases AI decision model 43 times faster than GPT-6 Luna

Kakao has unveiled a new AI model designed to support decision-making by AI agents, with faster processing speeds aimed at improving operational efficiency.

The company said Friday it released Kanana-2-v-decide, a multimodal AI model developed in-house, as open source on Hugging Face, the global AI platform.

The model was built to handle the rapid judgments AI agents must make while carrying out tasks. Unlike conventional large language models, which respond to user queries in full sentences, it delivers results in a structured format.

AI agents must make a continuous series of decisions — which tools to use, which information to reference — as they fulfill user requests. The model supports that process by supplying the judgment needed at each step so the agent can move quickly to the next action.

Kakao built the model on Kanana-2-v-3b, a multimodal language model it had previously developed. Because it can analyze text and images together, it can be applied to a range of tasks including customer inquiry classification, tool selection by AI agents and product image assessment.

The model analyzes input and returns results in three formats: probability, which indicates the likelihood that a given condition is true; choice, which identifies the most suitable option from a set of candidates; and score, a numerical value reflecting the probability at each evaluation stage.

Kakao also improved processing speed by applying a technique that represents input queries and answer choices as vectors — numerical groupings — and compares their similarity. On a single Nvidia H200 GPU, the model recorded an average response time of under 0.04 seconds, about 43 times faster than GPT-6 Luna.

The model also demonstrated competitive performance in benchmark evaluations. It ranked highest among all models tested on both Kanana Decision Bench — Kakao's proprietary evaluation dataset comprising 15,909 Korean-language questions — and Decision Index, a publicly available benchmark.

Models compared in the evaluation included Microsoft's Decision-1, Typesafe AI's Jev, Perplexity's pplx-decider-v1.1-27B and Upstage's Solar Decide.

Kakao applied its Kanana Open License to the model, permitting both research and commercial use. The company said the intent is to help researchers and businesses at home and abroad develop AI agent services.

Kakao plans to deploy the model first in internal projects, including web search and automated product category classification. It aims to refine performance in live service environments and, over the longer term, develop technology that integrates search and decision-making into a single model.

"This release is meaningful in that it broadens the foundation for applying cutting-edge AI technology to real services," said Choe Dong-jin, performance leader for applied AI models at Kakao. "We will continue to advance and open our technology so users can experience AI in their everyday lives."

Meanwhile, on Oct. 29 Kakao released two embedding models — Kanana-2-embedding — as open source. The models, which underpin personalized AI agent technology, analyze the semantic similarity between user queries and information to surface relevant results, with a focus on improving Korean-language search performance and processing efficiency.