Kakao 發佈開源多模態決策模型 Kanana-2-v-decide

Kakao has released its new AI model as open source. (Source: Kakao)

Kakao has unveiled an artificial intelligence (AI) model that analyzes text and images to classify customer inquiries or help AI agents select tools. Instead of generating sentence-based responses, the model provides decisions in the form of probabilities, selections and scores.

Kakao released its internally developed multimodal model, Kanana-2-v-decide, on Hugging Face on Nov. 6. The model is based on the company's proprietary multimodal language model, Kanana-2-v-3b, and accepts both text and image inputs.

The model provides three types of output. It can calculate the probability that a specific condition is true, assign probabilities to multiple options and identify the most suitable choice, or generate scores reflecting the probabilities of different evaluation levels. Developers can select the output format required for their services and connect the results to subsequent tasks.

For example, when a customer submits a photograph of a damaged product along with a replacement request, the model can analyze the image and inquiry to determine whether the product is damaged, classify the inquiry and assess the customer's emotional state. The service can then apply predefined processing rules to route the customer to a support channel or provide replacement instructions.

Kakao adopted an embedding architecture that converts inputs and candidate options into numerical vectors and compares their similarity. Vectors representing the same options can be stored and reused, reducing computational requirements and response latency when handling repeated requests.

The company said the model achieved an average response time of less than 0.04 seconds in tests using a single Nvidia H200 graphics processing unit (GPU).

The model is released under the Kanana Open License, which permits research and commercial use. Kakao plans to initially deploy it in internal projects, including web search and automated product category classification, to improve performance. The company subsequently intends to integrate search and decision-making capabilities into a single model.