KT will showcase an artificial intelligence (AI) model lineup specialized in the Korean language and culture later this year.
According to KT on the 25th, the company recently revealed its AI model concepts structured into four categories: proprietary models, collaborative models, open-source based models, and specialized models on its official website. Following the release of 'Mit:um 2.0' developed with its own technology in July, this is an attempt to actively expand the Korean language-centric AI ecosystem. Initially, it plans to launch 'GPT K' (tentative name) within the third quarter of this year.
This model, being developed in collaboration with Microsoft (MS), is based on OpenAI's GPT-4o and has specialized understanding of the Korean language and Korea's history, culture, and emotions by training it on proprietary data held by KT. It maintains the global top-level performance of GPT-4o while overall improving Korean language performance. KT explained that it refined Korean knowledge through its own data and fine-tuning, and applied a Korean specialized performance benchmark developed in collaboration with academia and industry.
An open-source based model 'Llama K' will also be released in the third quarter of this year. This version is based on Meta's 'Llama 3.x' but supplements the limitations of the original model's Korean capabilities with KT's data and learning know-how. It enhances the practical usability by strengthening unique Korean expressions such as honorifics and sentence-ending particles, as well as implementation capabilities for public sector formats.
Llama K will be available in two sizes: 11 billion parameters (11B) and 74 billion parameters (74B), capable of responding to general performance requirements and enterprise-level high-performance demands based on retrieval-augmented generation (RAG). Additionally, KT plans to introduce AI models specialized in specific professional fields, such as law, finance, education, and healthcare within the year.