[Edaily Reporter Lee So-Hyun ] Kakao(035720)has released an open-source multimodal decision-making model that rapidly processes the judgments and choices of Agent AI. Unlike general large language models (LLMs) that generate sentences, this model is designed to output results in the form of probabilities, choices, and scores, allowing Agents to quickly move on to the next task.

Kakao announced on the 6th that it has released its in-house developed multimodal model, “Kanana-2-v-decide,” as open source on Hugging Face.

As AI agents increasingly perform multi-step tasks that require them to repeatedly decide which tools to use, what information to reference, and what actions to take next, the development of models specialized in decision-making is on the rise.

Instead of responding to user questions with natural language sentences, Kanana-2-v-decide provides decision results in standardized formats such as △probability, △selection, and △scores.

The company explains that this allows services to immediately utilize the results for follow-up tasks, reducing the need to extract necessary conclusions from natural language responses.

Analyzing Text and Images Together to Determine the “Next Action”

Kakao developed this model based on its proprietary multimodal language model, “Kanana-2-v-3b.”

Since it can understand text and images together, it can be applied to various services, such as classifying customer inquiries, selecting tools for AI agents, and identifying product images.

For example, if a customer requests an exchange along with a photo of a damaged product, the model analyzes both the image and the inquiry to assess the likelihood of damage, the type of inquiry, and the customer’s emotional state using probabilities or scores.

The service can then use these results, combined with predefined criteria, to either connect the customer to a support agent or guide them through the exchange process.

The model employs an embedding-based architecture that converts input questions and answer choices into vectors—sets of numbers—to compare their similarity. Since vectors for identical answer choices are stored and can be reused, this reduces computational load and response latency for repeated requests.

Approximately 43 times faster than GPT-6

According to Kakao, Kanana-2-v-decide recorded an average response time of less than 0.04 seconds when running on a single NVIDIA H200 GPU.

The company explained that, based on its own measurements, the model is approximately 43 times faster than GPT-6 Luna.

In terms of decision-making performance, it also reported the highest performance among the compared models on both its proprietary evaluation dataset, “Kanana Decision Bench,” and the public evaluation dataset, “Decision Index.”

The comparison group included Microsoft’s “Decision-1,” TypeSafe AI’s “Jev,” Plexity’s “pplx-decider-v1.1-27B,” and Upstage’s “Solar Decide,” among others.

Application of an Open License for Commercial Use

Kakao has applied the “Kanana Open License” to this model, which permits not only research but also commercial use. The aim is to enable researchers and companies both domestically and internationally to utilize it for the development of Agent AI services.

Kakao plans to first apply the model to internal projects—such as web search and automatic product category classification—to improve its performance in real-world service environments. Going forward, the company will advance the technology toward integrating search and decision-making functions into a single model.

Choi Dong-jin, Performance Leader for Applied AI Models at Kakao, stated, “This is significant in that it broadens the foundation for applying the latest AI technology to actual services,” adding, “Through technological advancement and public release, we will provide AI experiences that users can tangibly feel in their daily lives.”