Kakao announced the release of its self-developed multimodal model, 'Kanana-2-v-Decide,' on the global open-source platform Hugging Face on November 6.
Unlike traditional large language models (LLMs) that respond to user inquiries with sentences, this decision-making model provides results in standardized formats such as probability (Noul), choice (Choice), and score (Score). This design allows for immediate application of results, reducing the need to extract conclusions from natural language responses. It is particularly useful for AI agents that perform multiple tasks based on user requests.
The new model is built on Kakao's own multimodal language model, 'Kanana-2-v-3b.' It understands both text and images, delivering results in three formats—probability, choice, and score—making it applicable for various services requiring judgment, such as classifying customer inquiries about damaged product exchanges or selecting AI agent tools and assessing product images.
Kakao implemented an embedding-based structure that represents questions and options as bundles of numbers, or 'vectors,' allowing for flexible comparisons of similarities. This approach reduces computational load and response delays by enabling the reuse of vectors for identical options, even with numerous classification categories.
The model achieves an average response speed of under 0.04 seconds with one H200 graphics processing unit (GPU), making it 43 times faster than GPT-6 Luna.
In evaluations using its own dataset, 'Kanana Decision Bench,' and the publicly available dataset 'Decision Index,' the model outperformed competitors, including Microsoft's 'Decision-1,' TypeSafe AI's 'Zeb,' Perplexity's 'pplx-decider-v1.1-27B,' and Upstage's 'Solar Decide.'
Kakao has applied the 'Kanana Open License' to this model, allowing for both research and commercial use, enabling domestic and international researchers and companies to develop various AI agent services.
The company plans to prioritize the application of this model in internal projects, such as web search and automatic product category classification, to continuously improve performance in real service environments. Kakao aims to enhance technology by integrating search and judgment into a single model.
Choi Dong-jin, leader of Kakao's Applied AI Model Performance, stated, "The release of this model is significant as it expands the foundation for utilizing the latest AI technology in real services. We will continue to provide AI experiences that users can feel in their daily lives through technological advancement and openness."
Meanwhile, Kakao is expanding the application range of its self-developed AI models. Following the open-source release of the lightweight language model 'Kanana Nano 2.1B' in February last year, it unveiled four small language models in July and two models, 'Kanana-2-Embedding,' that analyze user inquiries and document meanings to search for relevant information. With the addition of this decision-making model, Kakao is securing both information retrieval and judgment capabilities necessary for AI agents.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.